{"id":"113db53c-3595-4924-a93b-b1c327ab6cfa","arxiv_id":"2509.00010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CERA uses an autoencoder with latent-space alignment to learn climate-invariant representations, improving generalization of moist-physics parameterizations to a +4K climate without warmer-climate labels.","lead":"A new machine learning framework, CERA, learns climate-invariant representations from unlabeled warmer-climate inputs and improves predictions of atmospheric moist processes under +4C warming. It outperforms both raw-input and hand-crafted physics-based baselines without using any warmer-climate labels.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Marginal alignment via EMD does not establish the climate invariance of P(Y|Z) that the transfer step requires; this conditional invariance is assumed, untested, and central to the claimed generalization.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the EMD loss aligns only marginal latent distributions, while the downstream predictor requires climate-invariant conditional distributions P(Y|Z). This is not a manufactured concern; it is the logical gap between the method's objective and its claimed generalization. The empirical +4K R2 is positive evidence, and the ablation supports the role of alignment, so the paper is not without merit. However, the assumption is neither stated nor tested, and the absence of released code/data plus the hyperparameter selection described in the Supplementary Materials makes the reported improvement difficult to verify independently. These considerations justify the reader's CONDITIONAL verdict, but they do not push the verdict further: if the conditional-invariance test passes, the central claim stands. I therefore recommend no change to the verdict; the condition should be that this invariance be demonstrated directly and that code/data be released.","tokens_in":13800,"tokens_out":5718,"duration_ms":72069,"concrete_test":"Use the held-out +4K labels as an oracle (evaluation only, not for the submitted model): freeze the trained CERA encoder and hyperparameters, then train a second MLP predictor on +4K labels with the same architecture and loss. Compare its +4K R2 to CERA's reported 0.53. If the oracle predictor achieves materially higher R2 (e.g., approaching the control-climate R2 of 0.75), the gap is direct evidence that P(Y | Z_aligned) shifts with climate and that marginal alignment alone is insufficient. If the oracle R2 is close to 0.53, the conditional-invariance assumption is supported. As a cross-check, compare control and +4K residual distributions at matched values of Z_aligned using shared nearest-neighbor bins.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The transfer step rests on an invariance that the objective does not enforce. Equation (1) penalizes EMD(Z0, Z+4K), a marginal alignment of the latent codes, and the aligned channels are then fed to an MLP trained only on control-climate labels. For CERA's warm-climate skill to follow from the method, one additionally needs P(Y | Z_aligned) to be identical in the control and +4K climates. EMD gives no control over this conditional distribution: a latent space can have perfectly matched marginals while the target relation changes with warming. The paper never states or tests this assumption, and its own Section 4 concedes that 'not all physical processes can be cast into climate-invariant forms.' Moreover, the non-aligned latent channel is deliberately excluded from the predictor, so any climate-dependent information relevant to the tendencies is inaccessible to warm-climate predictions. The reported +4K R2 = 0.53 is empirical support that the assumption approximately holds for these outputs, but it is not a test of the mechanism: the hyperparameters were selected after 'initial tests on generalization performance' (Supplementary Materials), and code and data are not yet released. The central claim that CERA improves generalization without +4K labels is only as secure as this unverified conditional invariance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CERA, a two-stage machine learning framework for parameterizing moist-physics processes with improved generalization to a warmer climate. An autoencoder is trained on unlabeled inputs from both control (0K) and +4K aquaplanet simulations, with an Earth Mover's Distance (EMD) loss that aligns the per-level, per-channel marginal distributions of latent codes from the two climates. A predictor MLP is then trained only on control-climate labels, using the aligned latent channels as inputs. The authors compare CERA against a raw-input baseline, a physically informed RH+B baseline following Beucler et al. (2024), and an ablation without EMD alignment (CERA-noAlign), reporting R² scores for five subgrid tendencies in both climates. The main results are that CERA achieves a mean +4K R² of 0.53 versus 0.46 for RH+B and 0.26 for Baseline, that it captures the intensity distribution of precipitation including extremes better than the baselines, and that the EMD loss improves both accuracy and seed robustness. The paper also includes a direct comparison to Beucler et al. (2024) after retraining to include radiative heating, where CERA and RH+B are comparable in the +4K climate. The central claim is that CERA improves warm-climate generalization without using any +4K labels, by learning climate-invariant representations from data rather than hand-crafted features.","tokens_in":14054,"tokens_out":4143,"duration_ms":48094,"significance":"If the claimed result holds, CERA is a useful contribution: it offers a data-driven alternative to manual climate-invariant feature engineering, with a clean ablation showing that latent alignment matters, results averaged over five random seeds, and a meaningful set of offline diagnostics including vertical structure and precipitation extremes. The framework is potentially transferable to other parameterized processes and downscaling. However, the central claim rests on an invariance assumption that the objective does not enforce and the paper does not test: the predictor is trained on control labels, so transfer to +4K requires that the conditional distribution of the target tendencies given the aligned latent code be the same in both climates. The EMD loss in Eq. (1) aligns only marginal distributions of the latent codes. In addition, hyperparameters were selected after initial tests on +4K generalization performance, so the reported +4K R² values are partly selected rather than purely held-out. These concerns are fixable with additional diagnostics and clearer evaluation protocols, but they currently limit the strength of the generalization claim.","major_comments":[{"comment":"The transfer step rests on an untested conditional invariance assumption. The EMD term in Eq. (1) penalizes the divergence between the marginal distributions of Z0 and Z+4K, computed separately for each latent channel and vertical level, but the predictor requires P(Y|Z0) = P(Y|Z+4K) to transfer control-trained labels to the warm climate. The paper never states this assumption explicitly, and the EMD objective gives no control over it: two latent spaces can have perfectly matched marginals while the input-output relationship changes with warming. The empirical +4K R² of 0.53 is suggestive but not a test of the mechanism, particularly because hyperparameters were selected using +4K generalization performance. I recommend adding an explicit statement of the assumption and a diagnostic that uses a held-out subset of +4K labels (for evaluation only) to compare the conditional mean E[Y|Z] between climates, or a probe that measures how much a small correction head trained on warm labels improves the control-trained predictor.","section":"Section 2.2, Eq. (1)"},{"comment":"The hyperparameter selection procedure uses the +4K generalization performance, which conflates model selection with evaluation. The SI states that 'after initial tests on generalization performance, hyperparameters were selected through a sweep aimed at balancing latent alignment and predictive accuracy,' and the reported +4K R² values are computed on the same +4K data used for those initial tests. This means the headline mean R² = 0.53 for CERA is a selected performance, not an independent estimate, and the comparison with the baselines may be optimistic. I recommend a nested or three-way split (control training, warm validation for model selection, and a separate warm test for final reporting), or at minimum a sensitivity analysis showing that the ranking is stable across reasonable hyperparameter choices.","section":"Supplementary Materials, Training details"},{"comment":"The exclusion of the non-aligned latent channel from the predictor compounds the conditional-invariance concern. One of the three latent channels is deliberately excluded from both the alignment and the predictor, so any climate-dependent information in that channel that is relevant to the target tendencies is unavailable to warm-climate predictions. The paper reports that this partitioning slightly improves generalization for ice-related outputs, but it does not discuss the possibility that excluding a channel removes information that is needed precisely because P(Y|Z) changes with warming. The manuscript should at least acknowledge this trade-off and, ideally, test a variant that uses all channels in the predictor or a variant that aligns the conditional distributions rather than only the marginals.","section":"Section 2.2 and Section 3.1"}],"minor_comments":[{"comment":"There is a typo in the first paragraph: 'weather forcasting' should be 'weather forecasting.'","section":"Introduction"},{"comment":"The phrase 'with learning rates of 3 ×10−3 and and a weight decay' contains a duplicated 'and'; also the sentence 'EMD(Z0,Z +4K) decreases when Lreconstruction becomes smaller' is speculative and should be reworded or supported by an additional diagnostic.","section":"Supplementary Materials, Training details"},{"comment":"The approximation that the instantaneous surface precipitation rate excludes the surface ice sedimentation flux is described as 'typically small,' but this approximation is used for all precipitation diagnostics in Figures 5 and 6; it should be stated in the main text where the diagnostics are introduced, not only in the supporting information.","section":"Section 2.1 and Supplementary Materials"},{"comment":"The panels in Figure 2 use different vertical axis ranges, which makes cross-panel comparison of R² values difficult; I suggest either using a common axis where possible or adding explicit axis labels that make the differing ranges immediately visible.","section":"Figure 2"},{"comment":"The data and code are described as 'being prepared for public release' rather than released. For a paper whose main contribution is a method, providing the code and data at the time of review would substantially strengthen reproducibility; at minimum, the final version should contain permanent links.","section":"Open Research Section"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for JAMES and the empirical results are interesting, but the central generalization claim currently rests on an untested conditional invariance assumption, and the evaluation protocol includes model selection on the +4K test data. I would encourage the editor to require an explicit statement and diagnostic for the invariance assumption, a clarified evaluation protocol, and release of code/data before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: CERA is a solid, well-executed contribution to an important problem—getting ML parameterizations to generalize to warmer climates without labeled future data. It learns a climate-invariant representation by aligning latent distributions with EMD, and in offline tests it beats both a raw-input MLP and the hand-crafted RH+B baseline on +4K generalization. The ablation and five-seed averaging give the main result some spine, and the precipitation diagnostics are more informative than a single R2. This is the first application of learned latent alignment to moist-physics parameterization, as far as I know, and it is a reasonable alternative to manual feature engineering.\n\nWhat the paper does not do is prove the mechanism it claims. EMD aligns the marginal distributions of the latent codes across climates, but the transfer step needs the conditional distribution P(Y|Z) to be invariant. That is assumed, not tested. The stress-test note is right: marginal alignment alone does not guarantee conditional invariance, and the supplementary shows hyperparameters were selected after looking at generalization performance, which is mild target leakage. The reported +4K R2 of 0.53 is empirical support that the approach works, but it is not a test of the invariance assumption. Also, the evaluation is offline only; no online coupled test. Code and data are not yet released, so independent reproduction is not possible.\n\nThat said, these are not fatal flaws. The empirical result stands: CERA generalizes better than the baselines, and the ablation suggests the alignment loss contributes. The paper is honest about limitations, including that not all processes are climate-invariant. The hyperparameter selection is disclosed, albeit in the supplement. The main weakness is the gap between the method's stated mechanism and what the objective actually enforces.\n\nWho should read this: anyone working on ML parameterization or climate-invariant ML. It deserves a serious referee revision, not a desk reject. The revision should add a test of conditional invariance (e.g., compare P(Y|Z) between climates on held-out data), clarify the hyperparameter selection procedure, and ideally release code/data. I'd recommend sending it to review with major revision.","headline":"Learned latent alignment for climate-invariant parameterizations is a genuinely new idea and the offline results are convincing, but the paper's mechanism is under-specified and the hyperparameters were tuned with the warm-climate test in view.","tokens_in":14562,"tokens_out":2477,"would_cite":true,"duration_ms":26645,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that CERA, an autoencoder with explicit latent-space alignment, generalizes a moist-physics parameterization to a +4 K climate without using any labels from that climate, outperforming both raw-input and hand-engineered…","keywords":["climate-invariant representation","latent space alignment","Earth Mover's Distance","moist physics parameterization","deep learning generalization","out-of-distribution robustness","precipitation extremes","self-supervised learning"],"falsifier":"Take a held-out labeled +4 K sample, encode it with the trained CERA autoencoder, and compare for each vertical level the conditional expectation $E[Y \\mid Z]$ and its spread against the control climate's at matched latent codes; a substantial divergence would show that the transfer assumption is violated. A simpler version is to train CERA exactly as described, then fine-tune only the predictor on a small labeled +4 K subset and measure the jump in warm-climate $R^2$; a large jump would mean the control-trained predictor was out of calibration in the warm climate despite aligned latent marginals.","tokens_in":13604,"feed_emoji":"🌡️","tokens_out":5237,"duration_ms":53128,"temperature":0.7,"pith_summary":"This paper claims that a machine-learned parameterization of moist atmospheric physics can generalize to a climate 4 K warmer without ever seeing labeled training data from that climate. The trick is to make the model's internal representation, not its raw inputs, climate invariant: an autoencoder is trained on inputs from both the control and +4 K climates, and a distributional penalty pulls the two latent spaces together, while the downstream predictor is supervised only on control-climate outputs. Tested against a raw-input baseline and a hand-engineered relative-humidity-plus-buoyancy baseline, the framework (CERA) reports the highest mean $R^2$ in the warm climate (0.53 versus 0.46 and 0.26) and, unlike either baseline, tracks the enhanced frequency of extreme precipitation under warming. If this holds, it offers a route to parameterizations that do not depend on high-resolution labeled output from future climates.","feed_headline":"Without +4K labels, CERA beats physics-engineered baselines","feed_subtitle":"Latent alignment lets a control-trained model predict +4K moisture and precipitation extremes.","key_machinery":"The mechanism is a partitioned-latent autoencoder trained with a three-term loss: reconstruction on both climates, supervised prediction on control labels only, and an Earth Mover's Distance penalty between the control and +4 K latent distributions, computed per channel and vertical level as the distance between sorted samples. One of the three latent channels is deliberately excluded from alignment and prediction, giving the encoder a place to store climate-specific information that the decoder needs for reconstruction but that should not drive the predictor. Kernel-size-one convolutions transform each vertical level independently, so the climate-invariant encoding is a stack of per-level learned transformations rather than a vertically mixing feature extractor. The Earth Mover's Distance term is what forces the encoder to find structure common to both climates; the ablation shows that without it the warm-climate skill drops and is less stable across random seeds.","core_discovery":"On the paper's own terms, the central discovery is that explicit alignment of latent codes across climates is what transfers a control-climate-trained parameterization to a +4 K climate. Without any labels from the warmer simulation, CERA encodes vertical profiles of temperature and total water in both climates with shared convolutional weights, penalizes the Earth Mover's Distance between the two latent distributions, and predicts subgrid fluxes and tendencies from the aligned subset of the code. In offline evaluation, it attains mean $R^2 = 0.75$ (control) and $0.53$ (+4 K), ahead of the physics-informed RH+B baseline (0.72 and 0.46) and the raw-input baseline (0.72 and 0.26); it also reproduces the warming-induced shift toward heavier precipitation, most clearly above 300 mm/day, while the baselines over- or underestimate across the intensity range. The authors also report that ablating the alignment loss lowers accuracy and increases seed-to-seed variance, and that the remaining weak spot is the tropical boundary layer.","pith_inferences":["Editorial inference: A direct test of the conditional invariance assumption would sharpen the claim, since the Earth Mover's Distance loss aligns only the marginal distributions of latent codes; holding out a small labeled +4 K subset and comparing $E[Y \\mid Z]$ between climates would show whether the control-trained predictor remains calibrated in the warm climate.","Editorial inference: The kernel-size-one, vertically local design means vertical coupling between levels is currently ignored; adding non-local vertical context in the encoder might recover some of the boundary-layer skill the paper identifies as the main weakness.","Editorial inference: Because the experiments use only two uniformly warmed climates, whether the learned invariance extends to non-uniform warming patterns such as polar amplification is untested; training the alignment across several climates with varied forcing patterns would reveal whether the invariance is genuinely climatic or specific to a uniform +4 K sea-surface-temperature shift."],"forward_implications":["A moist-physics parameterization can be developed for the present-day climate and then applied to a warmer world, reducing the need for high-resolution labeled simulations in the target climate.","Machine-learned parameterizations need not rely on manually selected climate-invariant inputs; invariance can be learned self-supervised from unlabeled inputs of the target climate.","Precipitation intensity distributions, including the extreme tail, can shift correctly with warming in a model trained only on control-climate labels, at least in the tested aquaplanet setup.","The authors argue the same alignment recipe could transfer to other mapping tasks such as statistical downscaling, though those applications are not demonstrated here.","The remaining limitation is that boundary-layer skill degrades under warming, so not all physical processes transfer equally well through the aligned latent space."],"supporting_citations":[{"why":"Defines the hand-crafted climate-invariant inputs (relative humidity and plume buoyancy) that serve as the physically informed baseline and the comparison setup.","marker":"Beucler et al. (2024)"},{"why":"Supplies the neural-network parameterization architecture, the high-resolution simulation with coarse-graining, and the subgrid tendency outputs that CERA builds on and compares with.","marker":"Yuval et al. (2021)"},{"why":"Provides the +4 K warmer aquaplanet simulation used as the unlabeled target climate for testing generalization.","marker":"O'Gorman et al. (2021)"},{"why":"Introduces the Earth Mover's Distance used as the latent-space alignment penalty.","marker":"Rubner et al. (2000)"},{"why":"Gives the closed-form one-dimensional Earth Mover's Distance as the distance between sorted samples, enabling the efficient per-channel, per-level alignment loss.","marker":"Levina & Bickel (2001)"},{"why":"Documents the System for Atmospheric Modeling (SAM) used for the convection-resolving simulations that generate all training and evaluation data.","marker":"Khairoutdinov & Randall (2003)"},{"why":"Provides the pseudo-global-warming approach cited as a practical way to obtain warmer-climate inputs when real observations of the future are unavailable.","marker":"Schär et al. (1996)"}],"fun_headline_variants":["CERA aligns climates, then predicts +4K without labels","Latent alignment lifts +4K prediction without warm labels","Control-trained CERA masters +4K moisture and extremes","CERA beats physics features via latent space alignment","Zero +4K labels: CERA still tops physics baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole transfer rests on the assumption that, once the latent code is aligned, the relationship between that code and the moisture and energy tendencies is identical in the control and +4 K climates; the Earth Mover's Distance loss aligns only the distributions of the codes, so if warming changes that relationship, the control-trained predictor will be miscalibrated even when the latent marginals match perfectly.","fun_headline_variants_meta":{"raw":{"variants":["CERA aligns climates, then predicts +4K without labels","Latent alignment lifts +4K prediction without warm labels","Control-trained CERA masters +4K moisture and extremes","CERA beats physics features via latent space alignment","Zero +4K labels: CERA still tops physics baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2601,"prompt_tokens":1001,"completion_tokens":1600,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1518}},"tokens_in":617,"tokens_out":1600,"duration_ms":11145,"temperature":1.0,"reasoning_tokens":1518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:26:55.024494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out labeled +4 K sample, encode it with the trained CERA autoencoder, and compare for each vertical level the conditional expectation $E[Y \\mid Z]$ and its spread against the control climate's at matched latent codes; a substantial divergence would show that the transfer assumption is violated. A simpler version is to train CERA exactly as described, then fine-tune only the predictor on a small labeled +4 K subset and measure the jump in warm-climate $R^2$; a large jump would mean the control-trained predictor was out of calibration in the warm climate despite aligned latent marginals.","supporting_citations":[{"cited_title":", Gentine, P","cited_arxiv_id":null,"evidence_quote":"Defines the hand-crafted climate-invariant inputs (relative humidity and plume buoyancy) that serve as the physically informed baseline and the comparison setup."},{"cited_title":", O'Gorman, P A","cited_arxiv_id":null,"evidence_quote":"Supplies the neural-network parameterization architecture, the high-resolution simulation with coarse-graining, and the subgrid tendency outputs that CERA builds on and compares with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the +4 K warmer aquaplanet simulation used as the unlabeled target climate for testing generalization."},{"cited_title":"\\ Randall, D A","cited_arxiv_id":null,"evidence_quote":"Documents the System for Atmospheric Modeling (SAM) used for the convection-resolving simulations that generate all training and evaluation data."}],"review_version":1}