1. Introduction

In a public interview entitled dissection of the crisis of science, the statistician Andrea Saltelli shared his concerns over how easy it is to manipulate data or deceive with numbers (https://corbettreport.com/interview-1424-andrea-saltelli-on-the-crisis-of-science/). This matter had led him to extend the theory of sensitivity analysis to sensitivity auditing, aimed to assess the entire process of knowledge and model generation, including explicit or implicit assumptions, interests, stakes and motivations on part of the developers. It also drove his engagement with issues in the fields of epistemology, philosophy of science, and science for policy surrounding the responsible production of data.

Such concerns are particularly relevant when it comes to guide both mitigation and adaptation policies in the face of rapid climate change. The IPCC’s procedures establish comprehensiveness, objectivity, openness and transparency as guiding principles. A guidance note on consistent treatment of uncertainties is also available to help the IPCC lead authors, which has not been much modified since the fifth Assessment Report (AR5). As an AR6 coordinating lead author, I was well-positioned to observe that, despite the IPCC’s vigilance, such recommendations are however not always respected. In particular, they can be challenged by pragmatic lead authors, mainly focused on advancing the decisions they deemed essential to combating climate change and, apparently, unfamiliar with the issues raised by Andrea Saltelli. This does not invalidate the IPCC’s conclusions in any way, but it does raise interesting and important questions about the underlying philosophy and motivations.

In theory, the main objective of the IPCC first working group (WG1) is to provide a comprehensive summary of what is known about the drivers of climate change. An open and transparent review by experts and governments is an essential part of this process and should reflect a diverse range of views and expertise. Yet, the increasing technical complexity of climate sciences and the growing weight of digital ESM in this field seem to have overshadowed other forms of knowledge and has led to a fairly rigid and uniform view of what scientific priorities should be [1–3].

ESMs are coupled ocean-atmosphere general circulation models (GCMs) in which a representation of the carbon cycle (and other potential interactive components such as atmospheric chemistry or dynamic vegetation) is included. While they allow for the interactive calculation of CO2 concentration and thus the assessment of CO2 emissions compatible with a given level of global warming, this new degree of freedom largely explains the greater difficulty of CMIP6 models in simulating historical global warming [4] compared to the previous generation, which was essentially composed of GCMs. Although IPCC WG1 is generally expected to indicate where further research is needed, including for improving these models, several chapters within the AR6 did not really comply with this request. Climate research priorities are mostly discussed outside the IPCC, within various bodies of the World Climate Research Program (WCRP) where co-optation prevails over diversity. Such a dichotomy is understandable but can generate mutual misunderstandings between those who produce climate simulations and those responsible for their assessment.

In practice, some climate scientists are increasingly inclined to produce their own assessments, outside the scope of the IPCC [5]. Conversely, and most importantly, most IPCC authors do not participate directly in model development. Consequently, they often place more trust in climate projections than model designers and seem to underestimate the complexity of evaluating and calibrating comprehensive ESMs, where seemingly minor details or adjustments can have considerable, yet rarely publicized effects [e.g., [6–8]. This discrepancy echoes my previous concerns about the IPCC’s motivations. Is it to trigger decisions as much as to inform policymakers and leave public opinion to decide what is good or bad for itself and the environment? Does this lead them to provide increasingly overconfident assessments and, consciously or unconsciously, to fail to acknowledge certain important limitations in their understanding and quantification of climate change?

Such a debate has a long history. In 2008, the Alfred Wegener Medal was awarded to Pierre Morel for his leadership in the development of climate research. As a young scientist, I was bewildered by his lecture entitled “models and reality” (https://www.egu.eu/awards-medals/ alfred-wegener/2008/pierre-morel/). Nowadays, I better understand his warning about the fundamental difference between weather prediction models — routinely tested in operational forecasting systems — and global climate models (GCMs) that cannot be evaluated given the lack of perfect analogs of future climate. A year later, a prominent climate scientist in the US noticed that many people seem to ignore whether climate models are any good, and, even more importantly, what they are good for [9]. Unlike Pierre Morel, he regretted that climate models are seen at best as black boxes, and at worst as simply irrelevant to their detailed concerns, whereas they can help unlocking situations that are currently hindering policymakers. Yet, he acknowledged that such usefulness of GCMs was conditional upon appropriate use.

This crucial debate was reignited by Saltelli and Funtowicz [10] who argued that GCMs have become central to policymaking but that this reliance has outpaced the scientific community’s ability to ensure model quality. They called on scientists to restore confidence by clarifying why and where models are useful. Like Pierre Morel, they claimed that complex computational models appear as authoritative black boxes, obscuring the assumptions, parameters, and uncertainties that shape their outputs. This can give a misleading impression of certainty and erode science’s credibility when predictions fail. While such failures can hardly escape the IPCC assessments, the increasingly limited space given to model evaluation in recent reports is alarming in the face of the growing ESM difficulties to capture the early warning signs of climate change at the regional scale [11–15].

To help scientists improve their practices, Saltelli and Funtowicz [10] proposed a set of stringent quality criteria for responsible model development and use at the science-policy interface. At a time when some global modeling centres claim to be developing digital tools that meet the expectations of decision-makers for both mitigation and adaptation policies [e.g., 16–19], these concerns may deserve further discussion.

In the present opinion paper, I build upon my IPCC AR6 experience—as a coordinating lead author of working group one [20]—as well as on post-AR6 studies to argue that current ESMS are not the Achille’s heel of mitigation policies (Section 3). Moreover, the illusory search for perfect kilometer-scale (km-scale) ESMs (Section 4) seems at odds with the need to explore other sources of uncertainty in global climate projections, including scenario uncertainty (Section 2), modeling uncertainty (Section 5), and internal climate variability (Section 6). Alternative research priorities are briefly suggested, including a stronger focus on observations and on their use for constraining imperfect models (Section 7). The objective is not to close the debate but rather to identify its fault lines and apparent contradictions, in order to suggest a stronger diversity and a more rational use of computing resources within the global climate modeling community.

2. IPCC scenarios will never represent the true radiative forcing trajectory

Greenhouse gases (GHG) emission scenarios have always been an important integrating element in the IPCC assessments. Unlike idealized experiments (abrupt or gradual increase in atmospheric CO₂), they provide the opportunity to address policy-relevant questions: what is a consistent remaining carbon budget to keep global warming below a certain level and what would be a consistent route of action to achieve this goal, which impacts are unavoidable and which are reversible. Moreover, they support a coherent use of contrasted socio-conomic pathways across the three working groups of IPCC.

Yet, it has been argued that IPCC scenarios can be misused by climate scientists [21]. Symptoms of this misuse include the treatment of unrealistic extreme GHG concentration scenarios as equally plausible emission pathways and the comparison of climate projections across inconsistent global development trajectories. The claim that the IPCC has extended its mandate from literature assessment to literature coordination may appear excessive. Yet, the assertation that the careless use of scenarios in a vacuum of plausibility has led to the widespread production of myopic or misleading perspectives on future climate change and climate policy may be worth of further discussion.

It has been also argued that IPCC has paid too much attention to high-emission scenarios [22]. This criticism must be tempered by the very recent nature of the discrepancy that has occurred between the observed concentrations of GHG and the worst-case scenario considered by the IPCC. Moreover, the recommendation that the range between high-end and low-end scenarios should be reduced, so that climate impacts can be communicated more clearly, is at odds with the objective partioning of impact uncertainties [23–26]. Excluding the lowest emission scenarios would do little to reduce such uncertainties and would no longer allow IPCC to assess avoided impacts by pursuing more ambitious mitigation policies.

Model use in climate policy requires a clear description of underlying assumptions and framing choices [10]. Although it is common knowledge that IPCC scenarios are illustrative, they are often analysed as if they were reliable enough to justify the development of “hyperrealistic” ESMs [27] and as if the Earth system response was independent of the emission pathway [28]. In the AR6, the global mean surface temperature (GMST) change, or global warming level (GWL), has been proposed as an alternative dimension of integration for emission scenarios. Yet, the introductive chapter of WG1 warned that the transient and equilibrium states of certain GWLs can differ in their climate impacts and that a GWL does not uniquely define a change in global or regional climate state.

Lessons drawn from the IPCC AR6 have led to new recommendations regarding the role of scenarios in future assessments, including the enhanced integration of impacts into scenarios (...) and a stronger focus on near-term climate change [29]. Better integration can be obtained by including impacts directly in the development of the scenario narratives, thus improving the consistency across the three IPCC working groups. Yet, such an integration would be difficult to implement in practice since climate change mitigation is a global issue while impacts are felt at the regional scale. Moreover, the growing focus on GMST overshoot scenarios seem to be driven by theoretical scientific issues as much as by plausible socio-economic scenarios [30].

In the same vein, Sanderson et al. [18] claim that the current use of ESMs driven by concentration scenarios omits to explore the carbon cycle modeling as a dominant source of uncertainty. They propose to drive the next generation of ESMs directly with CO2 emissions, thus allowing computing resources to be allocated to better real-time information related to the detectability and verification of emissions reductions and their relationship to expected near-term climate impacts. Yet, emission scenarios represent a negligible source of uncertainty for near-term climate change [23] and the evaluation of mitigation policies is not possible within just one or two decades.

The former chair of the IPCC WG1 also called for careful attention to the equity and plausibility of scenarios underpinning future climate projections, with greater attention to biogeophysical constraints from a warming world – for instance, hard limits for sustainable use of groundwater and forest biomass at regional scales [31]. She argued that reactivity is also important in order to develop timely new robust knowledge exploring climate change and climate action implications. Such a suggestion is welcome but looks like a self-contradiction since time is usually needed to check the robustness of new findings or to assess the efficacy of climate policies.

Most governments and industries seem overconfident in the large-scale deployment of land-based carbon dioxide removal (CDR) techniques to avoid making the necessary drastic reductions in GHG emissions today [32–34]. This may not only lock us into a strong overshoot above the 1.5°C GWL, but it can also increase biodiversity loss [35]. CDR methods pose major technological, economic, and social feasibility challenges, threaten food security and human rights, and risk overstepping multiple planetary boundaries [31,36]. Therefore, all IPCC scenarios suitable with the Paris agreement rely on risky bets and questionable assumptions [37–38].

To sum up, IPCC scenarios must be considered as useful working hypotheses that should be as explicit as possible and openly debated. They encompass a wide range of plausible futures, which emphasizes the need of balanced strategies to reconcile economic development with the preservation of the environment, not just a stable climate [31]. An exclusive focus on radiative forcings seems to be diverting the attention of many scientists from other statistical, physical, ecological and societal dimensions of emission scenarios. These include practical aspects such as the possibility to distinguish the forced climate response from internal variability, but also socio-economic issues regarding their feasibility and sustainability. Unpredictable political drivers, not forgetting major volcanic eruptions, will always remain fundamental obstacles to anticipate the actual and long-term radiative forcing trajectory [39].

3. Current ESMs are not the Achilles’ heel of mitigation policies

Since the beginning of the 21st century, conventional coupled ocean-atmosphere GCMs, have been gradually replaced by ESMs which include a more or less sophisticated atmospheric chemistry and an explicit representation of the global carbon cycle. One of the key motivation is the possibility to simulate the life cycle of anthropogenic aerosols and to better assess the carbon-concentration and carbon-climate feedbacks [18,40]. The transient climate response to cumulative emissions (TCRE) of CO₂ was proposed as a key metric for guiding mitigation policies. Defined as the slope of the quasi-linear empirical relationship between the level of global warming and the cumulative emissions of CO₂, it shows a large intermodel spread [41] which limits our ability to predict how much carbon dioxide may be emitted before reaching a specific level of global warming.

The TCRE metric was however criticized for several reasons, the first of which is its reliance on correlations between cumulative time series [42]. It only refers to the cumulative emissions of carbon dioxide while non-CO₂ greenhouse gas emissions must be also cut rapidly to maintain a chance of complying with the Paris agreement. It is an emergent property of the multi-model ensemble rather than a physically-based concept, as indicated by the key role of water resources in the terrestrial carbon cycle [14,25,43]. Moreover, the TCRE was shown to depend on the emission pathway, not just total emissions [28]. Such criticisms do not totally invalidate the TCRE concept but they highlight serious limitations.

Another motivation behind the development of increasingly comprehensive ESMs is that the same stresses that affect terrestrial and marine ecosystems can be critical processes that determine the natural carbon sinks [18,44]. The call for km-scale atmospheric and ocean models, without concomitant scientific and computational investment in the land and biosphere, may however perpetuate the geophysical view of Earth and, thus, not provide the comprehensive actionable information needed for a changing climate [45]. The distinction between the climate change mitigation on the one hand, and vulnerability, impacts, and adaptation on the other hand is apparently overcome. The dream of an all-purpose model comes true [2,3], at least as an additional source of studies pointing out the fragility of current ESM projections [46,47]. Yet, these new ESM functionalities also raise a number of ethical issues [31] and call for commensurate responsabilities in the provision of reliable information to guide both mitigation and adaptation policies.

Splitting the climate change problem into two separate questions—what are the emission pathways consistent with a specific GWL on the one hand, and what are the consequences of such a GWL on the other hand—is potentially misleading since many manifestations of climate change— especially sea level rise, but also changes in the cryosphere and land surface hydrology—are emission-pathway dependent. It may suggest that an all-purpose climate model is not the most suitable option. The GWL framing indeed relies on the assumption that carbon cycle feedbacks do not represent a major source of uncertainty in climate projections at the regional scale, at least compared to other physical processes [48,49] and other well-known sources of uncertainty [23,25,72]. This hypothesis questions the use of comprehensive ESMs to deal with both mitigation and adaptation issues.

Even the most state-of-the-art ESMs are not necessarily mature enough to deal accurately with the carbon-climate feedbacks [e.g., 4, 33, 40].Yet, the on-going development of km-scale ESMs seems to be motivated by technical as much as by scientific reasons. The advent of exascale supercomputers and progress in numerical modeling has open the door to several international initiatives aimed to represent some atmospheric and oceanic processes more explicitly, using horizontal grid spacings equal to or finer than 10 km globally [e.g., [50]. Such km-scale GCMs were initially developed in parallel rather than jointly, with increasing model complexity. Yet, this is no more the case and these two research lines have recently merged, leading to the development of km-scale ESMs [51,52] (cf. Section 3).

Such a decision may be understood in the context of increasing climate impacts, including on terrestrial and marine ecosystems. Yet, it is not supported by other imperatives such as the need to assess and compare efficiently a wider portfolio of mitigation strategies [53]. As emphasized by the outline of the AR7, such strategies do not only encompass drastic reductions of GHG emissions but also include more or less efficient and safe solutions to remove—or compensate for—the residual emissions. While many related strategies—the so-called Carbon Dioxide Removal (CDR) methods—pertain to the biosphere [54,55], their assessment is mostly penalized by a lack of structural diversity in the representation of soil, vegetation and marine processes (e.g., soil respiration and photosynthesis) at much smaller scales than one kilometer [56–59].

CDR methods have been considered so far, including by IPCC, as more suitable and acceptable than alternative Solar Radiation Modification (SRM) techniques [60,61]. Yet, their potential and benefits have been challenged by many studies [32,33,62–64]. Moreover, SRM methods will deserve a much greater attention in the forthcoming AR7. Such a decision reminds us that IPCC was created by and for policymakers, potentially leading to difficult ethical issues not only for the IPCC lead authors but also for the whole climate modeling community. In this respect, it can be argued that ESMs can both help and hinder sustainable mitigation policies.

The IPCC AR6 and the Global Stocktake Report under the Paris Agreement both emphasized the need for more ambitious mitigation policies. However, a thorough synthesis of the most significant barriers to the adoption of such policies was still missing. According to a recent study based on machine-learning (ML), these barriers are not primarily related to scientific issues but involve struggles between political coalitions about the appropriate relative burden sharing [65]. While more research is needed on viable solutions for overcoming such socio-economic barriers, there is on the other hand sufficient evidence of the potential threat of both on-going and future climate change to trigger more ambitious mitigation policies.

Following the publication of the AR6, the relevance of maintaining the first IPCC working group (WGI) was questioned by some specialist journalists (cf. https://www.lemonde.fr/idees/article/2021/09/04/a-quoi-sert-encore-le-giec_6093419_3232.html). More than three decades after its creation, and as climate change continues to spiral out of control, it is legitimate to ask whether this group still serves any purpose. The press article argued that “the IPCC operates within an endless cycle. As soon as the 6th report is released, the 7th is launched. An 8h will follow, and so on. The expertise thus always appears incomplete, as if it were constantly being refined, perpetually inadequate (...) In reality, there is nothing in AR6 that would be likely to trigger political action, and which was not written in AR5. And AR7 will not say anything sufficiently different from what AR6 has established to enjoin action.”

Although I am convinced that the IPCC reports, including those of the WG1, are more necessary than ever, I fully agree that ESMs are clearly not the Achilles’ heel of current mitigation policies. The Paris Agreement was a political trade-off between—to be as brief as possible—Global South and Global North countries which had no well established scientific background. The IPCC mantra is that each increment of global warming matters and that there is no well-known GWL associated specific tipping points in the Earth system [e.g., 66–68]. In line with [10], climate modelers and the IPCC could thus make it more explicit that policy choices go beyond what models can predict. For instance, better quantifying the TCRE is not the key to improving mitigation policies. The ESM difficulty to agree in their estimates of the remaining carbon budget should not divert climate experts from other political and socio-economic issues [31,69,70] and climate scientists from other priorities [25,48,71–73]. In the remainder of this essay, I will argue that even the use of km-scale GCMs (rather than ESMs) is not fully justified within the framework of CMIP, and that adaptation policies should be designed as risk management strategies requiring, above all, better sampling of other sources of uncertainty than the current resolution of climate models.

4. Global climate models will never be digital twins of the Earth System

Although leading to increasingly complex and expensive numerical models [e.g., 10, 40, 74, 75], the improvement of ESMs and the increasing technicality of climate sciences has led to a parallel competition in the design of ultra-high resolution models [e.g., 16, 51, 52, 76, 77]. These so-called digital twins of the Earth system are being developed as advanced virtual models that accurately simulate the behavior of the Earth’s climate and ecosystems in real time [3]. Initially developed in the industrial sector, digital twins are generally used to optimize the operation and maintenance of physical assets, systems, and manufacturing processes. They may also constitute a training technology for the Industrial Internet of Things, where they can interact virtually with other machines and people. In this context, they are sometimes referred to as “cyber-objects” or “digital avatars.”

Presented as a technical and scientific revolution, such digital replicas of the Earth system do not only integrate a large diversity of physical and biogeochemical processes but also vast amounts of data from multiple sources to provide detailed simulations and predictions of climate patterns, environmental changes, and human impacts on the planet [3]. While the authors speculate about the values and ethics of climate science and the nature of climate research, they seem to ignore former recommendations made by there own community to ensure a judicious use of climate projections [e.g., 10,25,27,78,79].

Km-scale ESMs represent a natural outlet for the most powerful exascale supercomputers, at least before the emergence of GPU cards changed the game. The empirical Moore’s law (a doubling in transistor count every 18–24 months) has been upheld until recently. However, further increasing clock frequencies in smaller transistors has eventually resulted in power limitations. The way around this was to introduce parallelism in the processor itself, but these new processors require a partial or even total refactoring of complex scientific code [80]. This has led to calls for greater collaboration [17,81], including in Europe where most countries no longer have sufficient resources for developing and running such models. It may also increase the dependence of many countries on a few leading modelling centers for guiding their adaptation strategies.

Contrary to mainstream belief, digital twins are not simply replicas of an object at a specific point in time. They also encompass the process by which this object evolves from its conception to its final state. In other words, they reflect not only the object itself, but also its development and manufacturing history. As a result, the entire lifespan of this object must be considered, even after its final delivery. Such a definition suggests that this concept is hardly applicable to ESMs which rather look like constantly evolving numerical systems whose premises (i.e., the IPCC emission scenarios) are themselves subject to perpetual changes.

The development of km-scale operational ESMs could be better justified if we had the certainty that the involved institutes work for the common good and aim to fight against climate change. No doubt that this is however a naive perception of the current geostrategic issues surrounding climate modeling and the IPCC. Models may misuse their epistemic authority for a number of reasons, including occasion, opportunity and interest [82]. The development of digital twins has been motivated by the limited skills of current ESMs [81] but also by the rise of climate services and the common confusion between model resolution and model accuracy among decision makers. Although many modelling centres are currently developing such climate services as freely accessible products, it is easy to imagine that a global or regional leadership could lead to a more lucrative business model [83] or to more controversial applications [84].

Moreover, a susbtantial fraction of the climate modeling community still rely on regional climate models (RCMs) [85] and/or on increasingly performant statistical methods [86] to downscale global climate projections. In Chapter 10 of the AR6 WGI, Doblas-Reyes et al. [78] highlighted that “increasing resolution per se does not solve all performance limitations”. Early intercomparisons of convection-permitting RCMs suggest that significant model biases and a substantial inter-model spread can persist despite an explicit simulation of atmospheric convection [87–89]. Drawing more lessons from these RCMs could be a priority before engaging the climate modeling community in even more costly global simulations.

The claim that unresolved processes in the current-generation ESMs, such as oceanic eddies and atmospheric convective systems, hinder the simulation of climate change [17,90] has been contradicted by multiple studies. As an illustration, the inability of current ESMs to capture the observed tropical warming pattern may not be related to their coarse resolution [91,92]. Similarly, the double-ITCZ syndrom of many models is not necessarily improved at higher resolution [93]. Moreover, km-scale ESM simulations are often difficult to reconcile with an “honest communication of uncertainty.” As example, Moreno-Chamarro et al. [94] claimed to have used ensemble simulations to assess climate projections with various model configurations, but actually had a single high-resolution run to compare with previous ensembles.

Unlike what is commonly suggested, km-scale ESMs still need significant calibration efforts which are however made more difficult if not impossible by their huge numerical cost. Segura et al. [51] argue that a few days of km-scale simulations may be sufficient to build confidence on whether changes in parameter values converge to an improved climate response. Yet, they pretend to ignore that calibration is needed not only for fast atmospheric processes, but also at longer timescales in the ocean, land and biogeochemical ESM components. This needs perturbation runs of at least two years—rather than two days—in which climate drift can emerge from the noise due to internal variability [95].

Moreover, most climate modellers have not anticipated the rapid emergence of ML techniques in their field [96]. Arguably, the use of km-scale ESMs for operational applications, including the provision of fine-scale regional climate information in the framework of climate services, is increasingly challenged by the parallel development of regional climate emulators [86]. Similarly, the explicit simulation of deep convection in global atmospheric models, or of mesocale eddies in global oceanic models, is increasingly questionned by the possibility to develop less empirical parametrizations of small-scale processes via ML techniques [97]. The question of whether models based on deep learning and trained on atmospheric data can compete with weather models has already been settled [98]. Yet, climate models based solely on physical principles and the fundamental equations of motion may not be challenged by AI as long as future climate has no observational analogue [3,99]. Moreover, the evidence of hysteresis effects in ramp-up / ramp-down CO2 experiments [100] suggests that even the late climate response in overshoot scenarios may not be easily captured by ML techniques.

It is here further argued that physical understanding remains the cornerstone of IPCC Working Group 1, so that conventional ESMs will still play a key role for a long time to come. Far from being obsolete, they still make significant improvements through improved physical parametrizations [20,101] and improved evaluation strategies [102]. Moreover, their calibration is an essential element in the simulation and understanding of both present-day climate and climate change [103]. In particular, the use of perturbed parameter ensembles and of new statistical methods represent a potential breakthrough in the development and tuning of otherwise conventionel ESMs [104]. While they will never replace the need of structural diversity [105] regarding the simulation of both resolved and unresolved climate processes, such ensemble techniques suggest that increasing resolution may become an increasing obstacle for the ESM calibration.

To sum up, km-scale ESMs are symptomatic of an endless process motivated by strategic (leadership) and technical (exascale computers; 1 exaflop = 1 billion billion calculations per second) considerations as much as by scientific and societal needs. Digital twins of the Earth system will never replace the need for model diversity (Section 5) and for sampling the near-term climate change sensitivity to the intrinsic chaotic nature of the coupled ocean-atmosphere system (Section 6). While ESMs are crucial for representing carbon cycle feedbacks and better understanding how additional processes and new degrees of freedom may alter the response of the whole system under specific emission scenarios, such experiments do not justify the cost of digital twin simulations given their illustrative rather than predictive nature and other ESM limitations, not related to their current resolution.

5. Model diversity matters as much as model accuracy

Developing all-purpose ESMs at increasing spatial resolution belongs to the so-called unified approach of climate modeling which has been widely supported by WMO [106]. Such a strategy is however just one among several possible modeling strategies that have been proposed over the last two decades [96]. The “hierarchy approach” [107] argues for the development of hierarchies of models of lasting value, where the more complex models should relate in traceable ways to a simple model grounded in physical theory. The “pluralist approach” calls for a greater diversity in modeling efforts, including more attention to structural diversity, parametric uncertainty or more empirical modeling approaches [105,108].

The persistent resistance to the digital twin strategy [27,37,109] is partly grounded in the need for enhanced model diversity and the availability of improved calibration methods [104]. Digital twins have been often associated with risk, even in the industrial sector where they are better established and at a higher advanced level than in climate modeling [110]. Decision-making in the context of climate change fundamentally relies on probabilistic assessments. In this process, the ESM contribution should not be to add further risk but, on the contrary, to enable a comprehensive evaluation of major and sometimes irreducible sources of uncertainty via the possibility to run large enough ensembles at a reasonable cost.

Beyond the climate modeling community, the call for more diversity and interdisciplinarity is even stronger. Saltelli et al. [27] argue for instance that “the projects of digital twins do not engage with critical and interpretative social sciences” and contest the overall utility of the concept. They rather advocate for independent institutions to develop diverse models, prioritize communication with simple heuristic-based models, collect comprehensive data from various sources, and shift focus away from physics-centered variables to inform climate action. Such a strong position against km-scale models is often considered as a lack of overall confidence in scientific progress. Yet, such a progress could be less dependent on technogical developments and rather rely on well understood physical and statistical principles.

As an IPCC AR6 lead author, I was struck by the contrasted attention paid to modeling uncertainty between WG1 and WG2. As an illustration, projected water cycle changes were mostly mapped based on the ensemble mean model response within Chapter 8 of WG1, whereas the 5th and 95th percentiles of the model distribution were systematically shown in Chapter 4 of WG2. Global maps of local percentiles should not be confused with plausible global projections but are however extremely useful for adaptation policies which are designed at national to local scales, not at the global level. Many Chapter 8 lead authors were reluctant to openly communicate on modelling uncertainty and, thus, did not follow the recommended advanced IPCC approach (AR6 WG1 Cross-Chapter Box Atlas.1) regarding the use of stippling and hatching in global maps of ensemble mean climate change.

Model diversity encompasses both parametric and structural uncertainties. Given the increasing complexity of state-of-the-art ESMs, their number of empirical and unobservable parameters has increased significantly [111] and these parameters are often calibrated individually rather than in combination [6,103]. Model tuning is however increasingly recognized as a necessary component of developing and testing ESMs, which needs to become more transparent for both users of model output and other developers [7,112]. Knowing how ESMs are tuned is indeed essential to avoiding possible misattributions of skillful predictions. Practices differ significantly across modelling centres, including on key aspects such as the explicit use of historical transient records.

One key feature in this process has not much to do with unresolved processes due to the parametrization of atmospheric convection [17] but concerns the effective radiative forcing due to anthropogenic aerosol emissions. The direct perturbation of anthropogenic aerosols on Earth’s energy balance indeed remains highly model-dependent among the current-generation ESMs, primarily due to a poor calibration of aerosol optical properties such as the aerosol optical depth, the single scattering albedo, and the asymmetry factor [113]. Converging on such empirical parameters may however not lead to more consistent climate projections if no parallel efforts are devoted to the calibration of other fine-scale processes such as cloud radiative properties and cloud-aerosol interactions. In other words, model calibration is still in its infancy and may increasingly benefit from improved observations and methods [104,114].

Beyond parametric uncertainties, there is also a need for greater structural diversity in the way ESMs are being developed. It has long been recognized that perturbed parameter ensembles (PPE) with a single ESM tend towards insufficient dispersion and may be difficult to reconcile with observations [115]. Multi-model ensembles (MME) show a greater dispersion, suggesting that structural changes lead to a wider range of model behaviours than is obtained with PPEs. MMEs should thus represent a key component of uncertainty analysis. They must deserve a greater attention regarding the equivalent number of fully independent models [116]. The effective ensemble size is generally found to be much smaller than the actual number of models. This reduction is found to go beyond the problem of “same center” models. The lack of independence across models that share similar parametrizations or entire components may lead to overconfident climate projections [105,117,118].

Climates services rely on fine-scale climate information, hence the current use of both statistical and dynamical downscaling techniques before assessing the impacts of climate change [78,119]. The development of km-scale GCMs could be useful to avoid this more or less empirical step in the production of detailed projections at the territorial level and the difficult selection of the most relevant GCMs for driving current RCMs [120–121]. Yet, such a global-only modeling strategy will lead to a drastic reduction of the total number of ESMs. Beyond MMEs and PPEs, large initial condition ensembles [72,122] are also needed for a comprehensive sampling of plausible climate trajectories.

6. Internal variability matters as much as the forced climate response

The reproductibility of climate projections is fundamentally denied by the chaotic nature of the global climate system [23,72,123]. The true climate trajectory will never be simulated at the local scale given the key influence but limited predictability of internal climate variability. While the key objective of the climate modeling community should be to assess the statistical properties of the future climate’s attractor, the unreasoned development of increasingly expensive ESMs has led to an apparent dichotomy between ensemble decadal predictions and single run centennial climate projections [124].

It is now widely agreed that a too strong focus on mean climate may hide extreme weather events that can be even more harmful to ecosystems and human societies. As a results, it is also accepted that robust adaptation consists of choosing a set of strategies that have the greatest benefit across the broadest range of potential real-world outcomes [125–126]. Yet, there is still a WG1 inclination to focus on the forced climate response, regardless the expected large departures from this mean trajectory [72]. Moreover, the realization of single model initial condition large ensembles (SMILEs) has been promoted outside of CMIP.

Decision-makers will always have to deal with large irreducible uncertainties as a result of internal variability, regardless of the continuous improvement of climate models. Using one of the earliest SMILE, Fischer et al. [127] showed that trends towards more intense hot and less intense cold extremes may be masked or even reversed locally for the coming three to five decades. Likewise for precipitation extremes, multidecadal trends of opposite sign cannot be excluded over many land points. Similar conclusions have been obtained by subsequent studies regarding both temperature and precipitation extremes [123,128,129].

Such results have motivated the realization of other — around twelve — SMILEs over the past decade. Although not being part of CMIP, this is arguably the most prominent achievement of the climate modeling community since the release of the AR5. Such ensembles allow for a more robust separation between the forced model response and internal variability [23,130]. They also provide a useful test bed to benchmark more empirical statistical approaches devoted to this task. SMILEs are also particularly useful to estimate forced changes in variability, as expected from recent analyses and assessments of water cycle changes [20,131].

How large does a SMILE needs to be? This issue was addressed by Milinski et al. [132] using an objective method to estimate the required ensemble size that can be applied to any given application. The answer was obtained by splitting multi-millenial preindustrial control simulations. In some situations, five realizations are sufficient, while other applications require ensemble sizes well above 100 members. Larger ensembles are needed to quantify internal variability rather than the forced response. The largest ensemble size — typically more than 30 members — is needed to detect changes in internal variability [131] but remains more affordable than a factor 10 in horizontal resolution. This choice also depends on the acceptable error to the end-user and must be commensurate to the other potential sources of uncertainty.

SMILEs are also useful to address new scientific questions such as the emission pathway dependence of regional precipitation changes, the assessment of forced changes in highly variable atmospheric circulations, both in the troposphere and in the stratosphere, or the time-dependence of potential changes in the oceanic variability [133]. Based on a 100-member SMILE, this study was also used to investigate the ensemble size needed for a given problem and suggested that less than 100 members are sufficient in many cases, depending on the type of events and probabilities you are looking for.

The overall magnitude of the simulated internal variability was however shown to be model-dependent at the regional scale. Most CMIP6 models misrepresent the internal variability of surface temperature at the regional scale, thus potentially leading to spurious estimates of the frequency of extreme weather events [130]. While these results call for a better evaluation and calibration of internal variability in current ESMs, they may also suggest that further model participation to CMIP could require the achievement of typically 10–30 historical simulations. When the focus is on extremely rare events, dedicated ensemble boosting methods based on specific initial conditions can be also used as a surrogate for SMILEs [134].

Another attractive alternative to the use of SMILEs is to build narratives based on physically self-consistent past or plausible future events or pathways [71,135–138]. This so-called storyline approach does not need a priori probabilities but is rather focused on understanding the driving factors and their plausibility. Shepherd et al. [136] see at least four reasons for using this framing: improving risk awareness, strengthening decision-making, providing a physical basis for partitioning uncertainty, and exploring the boundaries of plausibility. Yet, the lack of probability may also lead to a considerable emphasis being placed on moral values in the conduct of adaptation policies, thus threatening the legitimacy and continuity of such policies.

Assessing climate change at specified warming levels rather than under different emission scenarios has also been proposed as an alternative method to provide robust estimates without running SMILEs. Yet, many responses of the Earth system and many impacts of climate change, especially at the regional scale, are not directly related to the level of global warming but rather depend on its spatial distribution [25,26]. Moreover, this framing may be no longer suitable for the AR7 given its particular focus on overshoot scenarios and solar radiation modification (SRM) remediation strategies. Evidence of complex hysteresis behavior in overshoot scenarios and the absence of a pervasive TCRE relationship in SRM experiments not only invalidate the GWL approach [18,139–141], but also call for larger SMILEs to be able to distinguish between increasingly detailed and subtle mitigation and remediation policies.

7. Conclusion

Climate science is currently at a crossroads [3]. The knowledge of the functioning of the climate system has increased substantially since the early IPCC assessments and new demands have arised from the needs of society to deal with on-going and future climate change. The development of regional climate information system, based on the most advanced global and/or regional ESMs, has however not led to highly consistent outcomes, either for a given emission scenario or for a specified global warming level. Further climate modeling progress is therefore needed to better inform the public about the local and remote impacts of climate change, and to guide decision-makers in optimizing their adaptation and mitigation agendas.

In this context, the rigour of the science underpinning the IPCC assessments and the development of climate services is increasingly important [79]. Demystifying climate projections and making model caveats more explicit can help both end users and policymakers to find better trade-offs between accuracy and reliability, as well as between overconfidence and false alarms. Buoyed by decades of relentless progress, climate modelers sometimes seem to have lost a critical perspective on their digital tools [27,142], and IPCC reports no longer pay much attention to their detailed assessment [20,29]. It may thus be both relevant and timely to reiterate the recommendations of Saltelli and Funtowicz [10] on why, where, and how climate models can play a decisive role in shaping climate policy.

To use ESMs in the design of climate policies, one should understand what is meant by the designer’s or policymaker’s knowledge base. In this process, one key information is the quantification of uncertainties and the understanding of their causes. An overconfidence in emission scenarios [30,143] and/or in ESM capabilities [144–146] may lead to inadequate mitigation strategies, just like an overconfidence in technology may lead to overestimate the remaining carbon budget which is compatible with the Paris agreement [147]. Yet, socio-economic and polical factors, rather than current ESM limitations, are by far the main obstacles to compliance with the Paris Agreement.

Similarly, natural hazards and climate change are not the only or even key drivers of recent climate-related disasters [148]. This observation should in no way lead to underestimating the increasing and harmful human influence on climate and weather hazards, but it calls for a wider perspective on exposure and vulnerability when dealing with adaptation. Moreover, the provision of fine-scale climate change information should not be done at the expense of a comprehensive assessment of related probabilities, including the contribution of parametric and structural modeling uncertainty, the influence and evolution of internal variability, and their potential interactions in driving unexpected tipping points. The multiplicative factor in computing resources that is needed to jump from current-generation to km-scale ESMs must account for a significant reduction in model integration time-step and is thus typcally around 1000, that is much more than the additional ressources that would be needed to run large PPE or SMILE at a standard horizontal resolution.

More than two decades ago, Dessai and Hulme [125] had already emphasized the challenge of attaching probabilities to future climate change. They warned that such probabilities are highly conditional upon the assumptions made, which need to be as explicit and transparent as possible. Their view on conditional probabilities was that climate scientists should not wait for perfect information before providing decision-makers with the best available scientific information for their questions. Arguably, the presence of irreducible sources of uncertainty in climate projections could even represent an additional lever for mitigation and adaptation policies rather than a pretext for inaction.

Overconfidence is a pervasive cognitive bias in subjective probability distributions because people fail to spread out probability mass to account for their own epistemic uncertainty [149]. Even if they are aware of this lack of knowledge, they do not always appropriately incorporate it into their judgement. This can lead to faulty assessments, unrealistic expectations and hazardous decisions. In particular, populations tend to be overconfident as long as benefits from contested resources are sufficiently large compared with the cost of competition [150]. This may explain why overconfidence remains prevalent even when it comes to fight against climate change.

Wrong models can prove very useful if used wisely [9,10,37,151]. Major breakthroughs in the field of science have been achieved after finding that existing theories and models failed to account for reality. There is no reason to believe that climate models are an exception to this rule, but neither is there sufficient evidence to suggest that the current ESM limitations stem primarily from a too coarse resolution [20,27,78]. Quite the contrary, the relentless promise of better-resolving and better performing models can foster the policymakers’ illusion that they would be wise to wait for yet another IPCC assessment before making decisions [152].

By recognizing that climate projections are not predictions, climate scientists can better support the development of no-regret adaptation strategies. They can also promote alternative research directions aiming at a better calibration and use of imperfect ESMs [104]. Related statistical methods can be further improved with the help of machine-learning [3] and could increasingly focus on climate sensitivity and recent climate trends [108,153]. Better understanding the key mechanisms underpinning climate change could also help to identify more robust emergent constraints on available projections [154–156].

Such recommendations are consistent with the outcomes of a recent review by young representants of the CMIP community [157]. In particular, they identified model evaluation strategies, model dependence, model selection and weighting methods, as well as uncertainty quantification as key issues in the use of MMEs. Although the diversity of research questions makes it impossible to create a set of universally transferable recommendations, they did highlight the growing relevance of statistics including ML techniques and SMILEs in shaping the future of climate ensemble analysis. They also advocated for awareness of the computational costs associated with climate modeling and analyses.

Such considerations do not mean that km-scale GCMs are not useful at all, but that their use could be restricted to the most suitable and relevant objectives. This means that the all-purpose model does not exist and that a “good model” (at least as good as possible) should not be confused with a high-resolution model. It is primarily a model that is suitable for addressing a particular question. Using the same model to deal with both mitigation and adaptation issues may not be the best option and spatial resolution may not be the main limitation of current ESMs [158–160]. In the first case, Calvin and Bond-Lamberty [158] suggest that more research and models are needed to robustly quantify the sign and magnitude of human-Earth system feedbacks. Yet, they also recognize that developing such integrated models entails significant complexity and cost, and that researchers should carefully assess the costs and benefits of doing so with respect to the object of study.

My suggestion is that the same philosophy should be applied to increased resolution. At what point does it become the most limiting factor in terms of a model’s use? What is its justification, and what is the most appropriate implementation in light of the objectives set? There are multiple evidence that increasing horizontal resolution is generally not sufficient for GCMs or RCMs to converge in their large-scale circulation or regional precipitation response to anthropogenic forcings [20,78,89]. Km-scale models are mostly and legitimately justified by an improved simulation of extreme weather events [17], but do not necessarily require global and/or transient experiments such as those which are regularly requested by CMIP [e.g., 89,161,162].

Regardless of model resolution, discrepancies between real-world signals and expectations based on standard ESMS are piling up with on-going climate change, including at the regional scale [153,163,164]. This has triggerred a loss of confidence in current GCMs [81] and increasing tensions regarding the future priorities of the global climate modelling community [104]. Philosophers of science characterize such situations as “crises.” Various solutions have been proposed such as developing a new ethic of climate modelling, using climate change signals to test assumptions, developing testable hypotheses, and revitalizing conceptual thinking by filling gaps across climate-system components and spatial scales [37,73].

Nevertheless, the increase in model resolution is still widely advocated as the most straightforward path forward [16,17,81,90]. As suggested in the introduction, it may be thus urgent to remember Saltelli’s warning about the use of mathematical models: “they are a great way to explore questions, but they are also a dangerous way to assert answers” [165]. Asking models for certainty or consensus is more a sign of the difficulties in making controversial decisions than it is a solution, and can invite ritualistic use of quantification. This is for instance obvious in the case of solar radiation modification which is increasingly seen as necessary to complement the current mitigation strategies. Yet, ethical considerations are here much more policy-relevant than our ability to quantify which level of global and regional warming could be thus avoided. Beyond a more thoughtful use of ESMs, a greater emphasis could be also put on improved quality-checked observations, including the development of longer and/or more reliable datasets [166–169]. A stronger focus on model-observation discrepancies could be also very useful to assess the reliability of climate projections [e.g., 13,14,15,153,170–173]. The Earth System modeling and Observations (ESMO) initiative from the World Climate Research Program (https://www.wcrp-esmo.org/) could be very helpful in setting such priorities.

Acknowledgments

This opinion paper expresses a personal viewpoint that in no way commits Météo-France. I would like to thank all my colleagues—most of whom are outside Météo-France—who helped stimulate these reflections, without animosity or prejudice toward a perspective that remains a minority view within the scientific community involved in climate modeling and climate change analysis.