{"id":"7c537e3c-9ad3-4a2a-a8d7-4fdbb7ee90e8","arxiv_id":"2506.18525","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Federated learning lets chemical companies train shared models on private data, and two case studies show it approaches centralized accuracy while outperforming isolated training.","lead":"This perspective tests whether competing chemical companies can jointly train machine learning models without sharing proprietary data. In two simulated case studies, federated learning matched centralized performance and beat locally trained models, suggesting a path to unlock private industrial data for ML.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case Study II's order-of-magnitude gain may be a source-data transfer effect, not a federated-learning effect; the paper's own heterogeneity caveat (Sec. 4.2.4) leaves the process-scale claim untested for structurally different plants.","rationale":"I read the paper as a perspective whose central claim is conditional: federated learning can give chemical companies near-centralized accuracy without sharing raw data, as demonstrated in two simulated case studies. The molecular case study is the stronger of the two: the even-random split reaches MSE 0.050 vs centralized 0.046, and the uneven-scaffold split still beats each client's local model despite non-iid data. I found no internal mathematical contradiction there. The process case study, however, carries the load for the 'processes' part of the title, and its result depends on source and target columns having nearly identical dynamics. The paper acknowledges this in Sec. 4.2.4, but the acknowledgment does not test the alternative explanation: because FedAvg weights by dataset size, the target's two trajectories can barely influence the global model, so the target's improvement may simply be receipt of a model trained on 768 similar-column trajectories. A source-only pretrained-model baseline would isolate whether federation, rather than transfer, drives the gain. I also note the lack of error bars and the post-hoc outlier exclusion in the full-data baseline, but I regard those as secondary to the transfer-confound concern. I agree with the reader's weakest_assumption and with the CONDITIONAL verdict: the paper should be read as promising proof-of-concept for near-identical processes, not as evidence that FL will transfer across structurally heterogeneous industrial plants. No verdict change is needed because the reader already made this conditionality explicit.","tokens_in":29152,"tokens_out":7574,"duration_ms":78607,"concrete_test":"Rerun the Sec. 4.2.2 scenario with a source-only pretrained baseline and with wider source heterogeneity, reporting all ten seeds, to test whether the FL gain is a transfer effect and whether it survives non-identical process dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FL yields near-centralized accuracy for process-scale system identification rests almost entirely on Case Study II (Sec. 4.2). There, the four source clients and the target differ only in the vapor flow rate V (1.6, 1.7, 1.8, 2.0 vs 1.9 kmol/s) of the same eight-tray methanol/propanol column, and Sec. 4.2.4 explicitly states that the resulting system responses 'show similar dynamics' (Fig. 8). Because FedAvg weights updates by dataset size (Sec. 2.1) and the target contributes only 2 of 770 trajectories, the global model sent to the target is essentially the source clients' model. The reported improvement (FL MSE 1.33e-4 vs private-data MSE 1.44e-3, RIPtoP 91%) is therefore consistent with a simple one-way transfer of a model trained on near-identical source columns, rather than with the collaborative privacy-preserving learning loop the paper advocates. The uneven-scaffold molecular case study does provide a non-iid test, but it still uses a single public thermodynamic dataset with shared molecular representations. As the paper itself notes in Sec. 3.2 and Sec. 4.2.4, real companies operate processes differing in configuration, control structure, feedstocks, and operating ranges; under such heterogeneity the demonstrated order-of-magnitude gain has no empirical support. The broader possibility that FL could help for processes may still be true, but the current evidence does not establish it beyond a conditional proof-of-concept.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This perspective paper argues that federated learning (FL) offers a way for chemical companies to jointly train machine learning models on proprietary data without disclosing the data. The authors review FL fundamentals, discuss potential ChemE applications, and present two case studies: (i) graph neural network prediction of infinite-dilution activity coefficients in binary mixtures under an even-random and an uneven-scaffold data partition, and (ii) Koopman autoencoder system identification of a distillation column, where one target company has only two trajectories and four source companies have 192 trajectories each. The results indicate that the federated global model approaches centralized performance in the molecular case and yields an order-of-magnitude error reduction for the target company in the process case. The central claim is that FL provides significantly higher accuracy than isolated training and can perform similarly to centralized training on combined data.","tokens_in":29400,"tokens_out":5847,"duration_ms":62769,"significance":"If the results hold, the paper provides a useful proof-of-concept for cross-silo FL in chemical engineering, complementing existing pharmaceutical examples with a molecular property prediction task and a process system identification task. The strengths are the open-source code and data repositories, the use of a realistic non-iid molecular scenario (uneven-scaffold with FedPer), and the clear reporting of the FedAvg workflow. The molecular case study is a credible demonstration that FL can approach centralized accuracy. The process case study is more limited than the abstract suggests, because the gain for the target company is largely attributable to near-identical source dynamics and is not isolated from simple one-way transfer. The paper's evidence thus supports a conditional proof-of-concept rather than the broad general claim in the abstract.","major_comments":[{"comment":"The process-scale FL result does not isolate the FL mechanism from one-way transfer. Because FedAvg weights updates by dataset size (Sec. 2.1, Eq. 1) and the target contributes only 2 of 770 trajectories, the global model is essentially a model trained on the four source clients' data. The reported MSE drop from 1.44e-3 to 1.33e-4 is therefore consistent with simply transferring any source-trained model to the target. Please add a baseline in which the target receives a model trained on one source client (or a simple average of source models) without federated communication rounds, and optionally a fine-tuned version on the target's 2 trajectories, to support the claim that the collaborative FL loop specifically provides the benefit.","section":"Sec. 4.2.4, Fig. 9"},{"comment":"The paper's own heterogeneity caveat undermines the broad abstract claim. The authors state that 'the system responses for different vapor flow rates show similar dynamics' (Fig. 8) and call more heterogeneous dynamics future work. Real industrial processes differ in configuration, control structure, feedstocks, and operating ranges, as the paper acknowledges in Sec. 3.2. The current evidence therefore does not support the general statement that FL yields significantly higher accuracy for process-scale system identification. Either temper the abstract and conclusion to a homogeneous-column proof-of-concept or add an experiment with structurally different columns (e.g., varying tray count, feed composition, or control structure).","section":"Sec. 4.2.4, final paragraph"},{"comment":"The abstract claims that FL 'can perform similarly to models trained on combined datasets from all companies,' but Case Study II never trains a centralized model on the union of all clients' data. The full-data baseline is 192 trajectories at V = 1.9 kmol/s, not a model trained on the combined 770 trajectories from all five companies. The 'combined datasets' claim is tested only in Case Study I. Please either evaluate a true centralized baseline on the union of all clients' data in Case Study II or restrict the corresponding claim to Case Study I.","section":"Sec. 4.2.2 and Sec. 4.2.4"},{"comment":"The 'each chemical company individually' part of the central claim is not evaluated for the four source clients. The paper reports performance only for the target company in Case Study II; there is no evidence that the source companies' models improve over their individual 192-trajectory models. Since the abstract claims that FL yields significantly higher accuracy than models trained by each chemical company individually, please report source-client performance or explicitly scope the claim to data-scarce target companies.","section":"Sec. 4.2.4"}],"minor_comments":[{"comment":"The word 'significantly' in the abstract is not supported by statistical significance tests. The authors report standard errors but do not show error bars in Figures 6 and 7, nor do they report confidence intervals or hypothesis tests for the observed differences. Either add significance tests or soften the wording to 'consistently higher' or 'clearly higher'.","section":"Abstract and Sec. 4.1.4"},{"comment":"The scaffold-split protocol is described only as partitioning by solvent scaffolds followed by a 70/15/15 split of each partition. If the 70/15/15 split is random, the same solvent scaffold may appear in both training and test sets, inflating absolute MSE estimates and the comparison to the centralized baseline. Please clarify whether a scaffold-aware split that keeps test scaffolds unseen was used, and if not, report a scaffold-leakage-free version.","section":"Sec. 4.1.2"},{"comment":"The caption contains the typo 'squarred' instead of 'squared'.","section":"Figure 9 caption"},{"comment":"The notation is slightly confusing: n_k is used both for the number of samples in client k and as the upper limit of the inner summation index i; consider using a different index or explicitly writing |D_k|.","section":"Eq. (1)"},{"comment":"The RIPtoP formula uses the convention that a perfect metric can be 0 or 1 depending on the metric, but the authors do not state the convention for MSE in the text; this is clarified in the case studies, but adding an explicit sentence would help readers applying the metric to other metrics such as RMSE or R^2.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"This is a perspective paper with illustrative case studies, so the experimental bar is lower than a methods paper. However, the abstract makes strong quantitative claims, and the process-scale case study is the main empirical support for the process-scale part of those claims. The key fix is to separate the FL effect from a simple transfer effect and to either broaden the heterogeneity experiment or narrow the claims. If the authors add the one-way transfer baseline and temper the abstract, the paper could become acceptable for the journal as a perspective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a perspective, not a benchmark paper. It argues that federated learning (FL) lets chemical companies train on proprietary data without sharing it, and it backs that with two small case studies. The first, GNNs for infinite-dilution activity coefficients, is the stronger piece of evidence: on an uneven scaffold split of a public dataset, FL gets close to centralized performance and clearly beats isolated training (global MSE 0.082 vs average client 0.076, RIPtoP 47%). The second, system identification of a distillation column with a Koopman/Wiener autoencoder, shows a large gain for a target client with only 2 trajectories (MSE 1.33e-4 vs 1.44e-3), but the four source clients and the target differ only in vapor flow rate, and the paper itself says the dynamics look similar (Sec. 4.2.4). That makes the result more one-way transfer from near-identical columns than a demonstration that FL handles heterogeneous industrial processes. The authors are upfront about this, so it is a limitation they own, not a hidden flaw.\n\nWhat is genuinely new is applying FedAvg/FedPer to these two ChemE tasks. Both case studies are described carefully: ten seeds, standard metrics, comparisons to local and centralized baselines, and a clear explanation of RIPtoP. Code and data are provided in GitLab repositories, though without pinned commit hashes, so reproducibility is conditional on repository state. The perspective sections are a useful synthesis of where FL could help in chemical engineering, and the references cover the relevant prior work.\n\nThe soft spots are real but not disqualifying. Statistical presentation is thin: no error bars or significance tests despite ten seeds, and the full-data baseline in Case Study II drops an outlier post hoc. The abstract's claim that FL yields \"significantly higher accuracy\" is supported for these settings, but the process-scale claim rests on a deliberately similar set of columns. Real plants differ in configuration, control structure, feedstocks, and operating ranges; the paper's own heterogeneity section says as much. So the broad promise is plausible but not established beyond a conditional proof-of-concept.\n\nI would send this to peer review. It is a clearly written perspective with reproducible demos, useful for the community, and honest about its scope. A referee should ask for error bars, pinned repository versions, and ideally a more heterogeneous process scenario, but none of that is fatal. I would cite it as evidence that FL is being actively explored in ChemE and would use the activity coefficient task as a baseline for future work.","headline":"A readable perspective with two honest proof-of-concept demos; the molecular case study is the stronger evidence, while the process-scale win is a transfer-like setting the authors themselves flag.","tokens_in":29984,"tokens_out":1746,"would_cite":true,"duration_ms":18605,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Chemical companies can jointly train ML models via federated learning without exposing proprietary data, and the resulting models beat local training and nearly match combined-data training.","keywords":["federated learning","chemical engineering","data silos","graph neural networks","activity coefficients","system identification","distillation column","data privacy"],"falsifier":"Re-run Case Study II with clients whose columns differ beyond vapor flow rate, such as different numbers of trays, feed stages, or mixtures, and check whether the target company's five-hour prediction MSE still drops from about $1.44\\times 10^{-3}$ toward $1.3\\times 10^{-4}$; if it stays near the private-data baseline, the similar-dynamics assumption is the reason the result held. For the molecular case, partition the activity-coefficient data by fully disjoint solute scaffolds; if the global model then no longer beats the data-richest client's local model, the federated benefit depends on scaffold overlap rather than on the federated mechanism itself.","tokens_in":28870,"feed_emoji":"🧪","tokens_out":6388,"duration_ms":62619,"temperature":0.7,"pith_summary":"The paper argues that chemical companies' proprietary data need not stay an obstacle to large-scale machine learning. It proposes federated learning, in which companies train models locally and share only model parameters, as the way to pool training signal while keeping data private. Two simulated collaborations support the claim: a graph-neural-network model for binary-mixture activity coefficients and an autoencoder-based model for distillation-column dynamics both improve markedly over company-private training. In the process case, federated training cuts the target company's prediction error by an order of magnitude, to roughly the level of a model trained on a full data set.","feed_headline":"Federated learning beats private-data ML in two chemistry cases","feed_subtitle":"Joint model training closes up to 91% of the gap to perfect prediction while keeping data on-site.","key_machinery":"The load-bearing mechanism is FedAvg (Federated Averaging): a server broadcasts the global model to clients, each client trains it on local data, and parameters are aggregated by weighted averaging across communication rounds, so training signal is pooled without data leaving the company. The two case studies embed FedAvg in, respectively, a graph neural network that forms a mixture fingerprint from solute and solvent molecular graphs before predicting activity coefficients, and a Koopman-style Wiener model that sandwiches a linear state-space model between an encoder and decoder to forecast distillation-column dynamics. The RIPtoP metric (relative improvement of proximity to perfection) converts raw errors into the fraction of the gap to a perfect model that federated learning closes, which is how the paper quantifies benefit.","core_discovery":"On the paper's own terms, the central claim is that ML models jointly trained with federated learning are significantly more accurate than models trained by each chemical company on its private data alone, and can perform nearly as well as models trained on a combined, centralized data set. In the even-random molecular case, the federated global model reaches a test MSE of $0.050$, close to the centralized model's $0.046$ and below the $0.064$ average of the individually trained clients. In the uneven-scaffold case, where clients keep their prediction heads private and share only the embedding layers, federated training still closes 47% of the gap between the local-model baseline and a perfect model. In the process-scale case, a company with only two trajectories reaches a multi-step prediction MSE of $1.33\\times 10^{-4}$ through federated training, versus $1.44\\times 10^{-3}$ when training on its own data, corresponding to a 91% RIPtoP.","pith_inferences":["I would expect the strongest industrial payoff where data heterogeneity is moderate and the shared input structure is standard, such as common unit operations run at different operating points, because the process case study only demonstrates transfer across vapor flow rates.","If federated models generalize this well, the economic value of a company's data may shift from exclusive custody toward contribution to a federation, which would make data valuation and fair incentive design central business questions.","A testable extension is to repeat the molecular case with strictly disjoint chemical spaces across clients; if the global model's advantage shrinks, the benefit depends on overlap in molecular scaffold space rather than on the federated mechanism alone.","Adversarial model extraction, already studied for drug-discovery federated learning, becomes more serious at process scale because shared model parameters may encode operating conditions; the paper lists this as an open research direction."],"forward_implications":["If the case-study results hold, chemical companies can jointly train predictive models on proprietary data sets without shipping data, so data scarcity ceases to be a reason to forgo machine learning.","Partially shared models, where only embedding layers are shared and prediction heads stay private, still deliver most of the benefit, giving a privacy-preserving middle ground for collaboration.","A company with very scarce process data can reach accuracy comparable to full-data training by joining a federation of similar processes, as the distillation case shows.","The same federated workflow is transferable to flowsheet digitization, anomaly detection, process design, and other chemical-engineering tasks where data is locked in silos.","Federated training offers a practical stand-in for the infeasible centralized ideal, since aggregating proprietary data across competitors is not a real option."],"supporting_citations":[{"why":"Defines federated averaging (FedAvg), the aggregation algorithm used in both case studies.","marker":"[29]"},{"why":"Supplies the formal federated-learning setting, including client distributions and the cross-silo interpretation the paper adopts.","marker":"[30]"},{"why":"Provides the large-scale pharmaceutical federation precedent and the partial-model-sharing design used in the uneven-scaffold scenario.","marker":"[43]"},{"why":"Provides the sanitized infinite-dilution activity coefficient dataset used in Case Study I.","marker":"[151]"},{"why":"Supplies the mixture fingerprint graph-neural-network architecture for binary activity coefficients.","marker":"[194]"},{"why":"Original activity coefficient dataset from which the federated partitions are drawn.","marker":"[196]"},{"why":"Software framework that implements the federated communication and aggregation loop.","marker":"[201]"},{"why":"Supplies the Wiener-type Koopman autoencoder model and distillation column simulator for Case Study II.","marker":"[204]"}],"fun_headline_variants":["Federated learning lifts chemistry AI without exposing data","Joint model training nears central-data accuracy in chemistry","Private data, shared gains: federated learning for chemistry","Federated learning beats siloed training in chemical cases","Chemical firms can train better ML jointly, data stays local"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The process-scale case study assumes that five distillation columns differing only in vapor flow rate have sufficiently similar dynamics that one global model trained on $V$ values of 1.6, 1.7, 1.8, and 2.0 kmol/s transfers to $V=1.9$ kmol/s; if real companies run structurally different processes, the demonstrated gain may not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Federated learning lifts chemistry AI without exposing data","Joint model training nears central-data accuracy in chemistry","Private data, shared gains: federated learning for chemistry","Federated learning beats siloed training in chemical cases","Chemical firms can train better ML jointly, data stays local"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1355,"prompt_tokens":953,"completion_tokens":402,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":323}},"tokens_in":569,"tokens_out":402,"duration_ms":4364,"temperature":1.0,"reasoning_tokens":323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:47:33.559748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Case Study II with clients whose columns differ beyond vapor flow rate, such as different numbers of trays, feed stages, or mixtures, and check whether the target company's five-hour prediction MSE still drops from about $1.44\\times 10^{-3}$ toward $1.3\\times 10^{-4}$; if it stays near the private-data baseline, the similar-dynamics assumption is the reason the result held. For the molecular case, partition the activity-coefficient data by fully disjoint solute scaffolds; if the global model then no longer beats the data-richest client's local model, the federated benefit depends on scaffold overlap rather than on the federated mechanism itself.","supporting_citations":[{"cited_title":"A smile is all you need: predicting limiting activity coefficients from SMILES with natural language processing","cited_arxiv_id":null,"evidence_quote":"Provides the sanitized infinite-dilution activity coefficient dataset used in Case Study I."},{"cited_title":"Rittig, Kobi C","cited_arxiv_id":null,"evidence_quote":"Supplies the mixture fingerprint graph-neural-network architecture for binary activity coefficients."},{"cited_title":"Model performances evaluated for infinite dilution activity coefficients prediction at 298.15 k","cited_arxiv_id":null,"evidence_quote":"Original activity coefficient dataset from which the federated partitions are drawn."},{"cited_title":"Schulze, Danimir T","cited_arxiv_id":null,"evidence_quote":"Supplies the Wiener-type Koopman autoencoder model and distillation column simulator for Case Study II."}],"review_version":2}