January was the month chemical AI stopped grading itself on a curve. Researchers spent it building benchmarks designed to catch their own tools cheating, and the tools obliged. Group contribution models turned out to have been quietly grading their own homework for decades. Twelve chemical foundation models were asked to predict something genuinely new, and twelve of them declined. The field called this maturing. It looked a lot like a bad performance review.
Researchers published AutoSID in Industrial & Engineering Chemistry Research, an automated system identification framework built on Machine Learning Operations principles. The point was to remove the human from model selection. Engineers picking a model architecture tend to pick the one they already know, which makes every published comparison a little suspect. AutoSID automates the whole lifecycle instead, running feature engineering, hyperparameter tuning, structure selection, and validation without an investigator's thumb on the scale.
The pipeline searches model structures using Tree-structured Parzen Estimators with K-fold cross-validation. On large data streams where cross-validation gets expensive, it switches to Akaike and Bayesian information criteria to pick a parsimonious structure instead. It evaluated 12 model classes across 11 process systems engineering case studies, covering continuous stirred-tank reactors, distillation columns, four-tank systems, and industrial bioreactors.
Tree ensembles won. XGBoost and LightGBM beat both classical linear models and deep recurrent networks on test-set error, and cut training time by up to an order of magnitude against the deep learning entries. The practical read for a process engineer is that overparameterized neural networks are usually unnecessary for dynamic modeling and model predictive control surrogates.
The results came from benchmark synthetic and semi-industrial plant data. Live plants bring non-stationary noise, sensor drift, and missing-data events that these case studies did not test. Still, the field spent several years assuming bigger networks meant better process models, and January's most rigorous benchmark handed the trophy to gradient-boosted trees.
A second Industrial & Engineering Chemistry Research paper introduced MGFF, short for Multi-Modal and Multi-Granularity Feature Fusion. Molecular property models usually pick one representation. They read a molecule as a SMILES string, or they read it as a graph. Each choice creates a blind spot the other would have covered.
MGFF runs both encoders at once. A BERT-based sequence encoder trains on SMILES strings through three simultaneous pre-training tasks, predicting graph-level atomic features, MACCS fingerprints, and sequence tokens. A hierarchical graph encoder handles topology, combining a 78-dimensional atomic feature vector covering atom type, valency, aromaticity, and hydrogen count with motif-level substructures. The two feature sets then merge through Low-Rank Multimodal Fusion, which computes cross-products between modalities using low-rank tensor decomposition so the combination does not become computationally ruinous.
The numbers held up across four standard benchmarks. MGFF reached an RMSE of 1.027 on SAMPL against 1.103 for the MMFDL baseline, 0.547 on Lipo against 0.725, and 0.576 on Delaney against 0.620. On the BACE bioactivity classification task it hit an AUC-ROC of 0.906, against 0.890 for the HiGNN baseline.
The cost is memory. Running two encoders and a tensor fusion layer takes considerably more allocation during training than a single GNN or a plain SMILES transformer. Reading a molecule twice turns out to work better than reading it once, at roughly the price you would expect.
A ChemRxiv preprint, not yet peer reviewed, went after a method chemical engineers have trusted for decades. Group contribution models estimate physical properties by summing contributions from a molecule's functional groups. Their published accuracy numbers look excellent. The preprint argues those numbers were inflated the whole time, because the models were fitted on all available data and then evaluated on molecules whose functional groups already appeared in training.
The fix is a splitting framework combining Butina clustering on molecular fingerprints with explicit functional group partitioning. It enforces two things at once. Every functional group must appear in the training set, so no parameter is left undefined, and chemical diversity must stay separated across training, validation, and test splits. The authors then re-ran traditional group contribution methods, hybrid machine-learning group contribution models, and 12 machine learning algorithms across 739 organic compounds, targeting critical temperature, critical pressure, critical volume, and acentric factor.
Group contribution accuracy collapsed under the leak-free split. Graph neural networks held their performance, posting lower RMSE and MAE on external test sets than both the classical models and the hybrid architectures.
This is a preprint making a strong claim against a large body of published work, and the dataset covers 739 organic molecules with common heteroatoms and halogens. Organometallic complexes, ionic liquids, and macromolecules were excluded. The stakes are not academic, though. Property estimates feed relief valve sizing, distillation column design, and reactor heat exchanger specifications, so a critical pressure estimate that is quietly worse than advertised is a plant safety input, not a leaderboard entry.
The BOOM benchmark, released as a preprint, tested 12 leading chemical foundation models across 10 out-of-distribution extrapolation tasks. The test sets were partitioned to hold property values and chemical scaffolds outside the training distribution, which is a deliberately unkind setup and also the only one that matters.
Every model degraded severely. ModernBERT narrowed the gap against standard transformers somewhat, but no model in the set generalized well out of distribution across all tasks. The benchmark compared them against domain-specific GNNs including DimeNet++ and SchNet, and against traditional supervised models.
The problem is structural rather than embarrassing. Materials discovery is out-of-distribution work by definition. Finding a material with properties better than anything known requires predicting outside the range of everything the model has seen. A model that interpolates beautifully within known chemical space and falls apart outside it is well suited to a task nobody needed done.
The benchmark covered small molecules, so biological macromolecular extrapolation stays unmeasured. The governance point is worth stating plainly: deploying AI-designed materials into batteries or structural alloys without accounting for out-of-distribution failure risks material fatigue and unexpected structural failure.
Running alongside BOOM in the same month, and covered in the same preprint entry, Crystalyse took the opposite posture. It is an agentic materials design platform built on reasoning models, using OpenAI o3 and o4-mini through the OpenAI Agents SDK. A central reasoning model orchestrates crystal structure generators, fast stability filters, and DFT validation engines through Model Context Protocol connections, an open standard that lets a model discover and call external tools without bespoke integration code.
It runs in three modes, Creative, Rigorous, and Adaptive, which trade search breadth against validation depth. In creative mode it screened complex quaternary oxide structures and sodium-ion battery cathode candidates in under 120 seconds per query.
Crystalyse depends entirely on the tools underneath it. A wrong surrogate output or a bad fast-DFT approximation propagates straight through the agent's reasoning tree without anything flagging it. Put the two January materials stories side by side and the shape of the field emerges: the agents got dramatically faster at asking questions in the same month the benchmarks confirmed the models cannot reliably answer the new ones.
LARC, released as a preprint, addresses a real failure in automated synthesis planning. Retrosynthesis models propose routes to a target molecule, and those routes regularly run through precursors that are carcinogenic, pyrophoric, controlled, or simply unbuyable. Filtering afterward is expensive and often leaves zero valid routes standing.
LARC splits the job across two agents. A Synthesizer module builds retrosynthetic trees from the target down to available building blocks. An Evaluator module, which the authors call Agent-as-a-Judge, audits each proposed intermediate and reaction step against user constraints, querying cheminformatics tools and safety databases through natural language tool calls. Its feedback prunes prohibited branches during the search rather than after it.
Across 48 constrained retrosynthetic tasks spanning safety hazards, cost limits, and structural prohibitions, LARC hit a 72.9% success rate, well ahead of unconstrained LLM baselines and post-hoc re-ranking pipelines.
The guardrail is only as complete as the databases behind it. If an API fails to flag a novel toxic compound, the Evaluator passes the route. That matters more here than in most benchmark caveats, because the safety case for this system is preventing autonomous labs from planning routes to chemical weapons precursors and explosives. A judge that has not heard of a hazard will not object to it.
Chemical engineers at the University of Wisconsin-Madison and the University at Buffalo published a computational framework, as a ChemRxiv preprint, for designing Solvent-Targeted Recovery and Precipitation processes. Multilayer plastic films are everywhere in food and medical packaging, and they combine incompatible polymer layers that mechanical recycling cannot separate. The usual fate is incineration or landfill.
The framework operates across four connected scales. Molecular thermodynamic models predict temperature-dependent polymer solubilities across candidate solvents. Process simulation models the dissolution, filtration, and precipitation unit operations. Techno-economic analysis calculates capital cost, operating cost, energy duties, and a minimum selling price per kilogram of recycled resin. Life-cycle assessment quantifies global warming potential against virgin resin production.
Previously this work required iterative physical experimentation to find a workable dissolution sequence. The framework screens waste streams containing dozens of distinct polymer layers in seconds, returning solvent choices, operating temperatures, recovery yields, and net CO2-equivalent savings. Solvent selection also prioritizes non-toxic, low-VOC options, which keeps the recycling process from becoming its own environmental problem.
Validation rests on clean thermodynamic solubility models. Real post-consumer waste arrives with dirt, pigments, adhesives, and degraded polymers, none of which appear in a clean solubility calculation.
The Journal of Chemical Education published paired findings in January that land in opposite directions. Pre-service chemistry teachers using generative AI showed measurable gains in Technological Pedagogical Content Knowledge, the framework describing how teachers combine subject knowledge with teaching method and technology. The gains concentrated in translating concepts across Johnstone's triangle, the macroscopic, symbolic, and particulate levels chemistry students are expected to move between. Teachers generated particulate-level animations, prompt-guided lesson structures, and diagnostic assessments, and their TPACK scores went up against an unassisted lesson-planning baseline.
The companion safety audit queried ChatGPT, Copilot, and Gemini across laboratory safety scenarios covering chemical compatibility, fire hazards, toxic gas ventilation, and spill mitigation, then scored the answers against OSHA, NIOSH, and ACS standards. The models produced plausible and dangerously wrong procedures in multiple instances, including improper chemical storage recommendations and hallucinated safety equipment protocols.
Both results are real and they do not cancel out. Generative AI is a capable lesson-planning aid and an unsafe unassisted laboratory safety consultant, and chemistry departments need policies that say both sentences. The TPACK cohort was small and focused, so long-term student outcomes need longitudinal work.
The dossier records this bluntly. A tool good enough to explain a concept to a class is not thereby good enough to tell that class which chemicals to store next to each other.
January's throughline was self-examination, and the field did not enjoy the results. A decades-old property estimation method lost most of its reported accuracy to a better data split. Twelve foundation models failed the one task materials discovery actually requires. Deep learning lost a process control benchmark to a decision tree. None of this is bad news. A discipline that can catch its own tools overstating themselves is in better shape than one still publishing the inflated numbers, even if the month reads like an audit report with a chemistry degree.
Questions about this issue or the underlying research
info@cheairesearch.com