November's most useful finding was a negative one. A self-driving lab optimizing a carbon capture electrolyzer found that peak reactor yield sits at neither maximum current density nor maximum CO2 utilization, which means every team that tuned those variables one at a time has been optimizing toward the wrong place. ExxonMobil published a related argument from industry, that purely empirical models fail on scale-up. The month closed 2025 on the theme it kept returning to: the constraint is what makes the model useful.
Researchers in the Berlinguette Group at the University of British Columbia ran closed-loop Bayesian optimization on an operating reactive CO2 electrolyzer using their ADA self-driving laboratory platform. The work is a ChemRxiv preprint.
Reactive carbon capture combines capture and electrochemical conversion in one unit, skipping the energy-intensive thermal desorber that conventional capture requires. That makes it attractive and hard to tune, because the variables interact through ionic strength, species solubility, and local mass transport at the gas-diffusion electrode.
The platform optimized across a six-dimensional space covering bicarbonate and carbonate concentrations, total liquid flow rate, gas feed rate, applied current density, and reactor temperature. A Gaussian Process surrogate with an Expected Improvement acquisition function navigated the trade-offs, running on live hardware without human intervention.
The result is the interesting part. Peak reactor yield occurred at a CO partial current density below the single-variable maximum, and at a CO2 utilization rate below its single-variable maximum. Neither individual metric was maximized at the point where the reactor actually performed best. Traditional step-testing tunes parameters sequentially and structurally cannot find that point, because it never looks at a setting where two variables are simultaneously not at their own best values.
The work covers a single bench-scale flow cell. Membrane degradation over time, impurity accumulation in recycled eluents, and stack-scale hydrodynamic non-uniformity were all outside the optimization space, and all three are what determine whether the finding survives contact with a commercial unit.
The safety architecture deserves attention. The team released the dataset, control logic, and hardware designs open-source, and they specify that automated pressurized liquid-gas electrochemical systems need hardware-level interlocks, automated pressure relief and gas leak detection, operating independently of the AI optimization loop. An optimizer exploring a six-dimensional space will eventually propose a combination nobody anticipated, and the interlock that stops it cannot be a parameter the optimizer controls.
A team at Dalian University of Technology published a generative screening framework in Environmental Science & Technology, peer reviewed, targeting per- and polyfluoroalkyl substances.
The problem is structural. Non-target screening matches mass spectrometry features against reference databases like PubChem or EPA CompTox, which only contain compounds someone already catalogued. Industrial PFAS substitutes and degradation products are frequently not in there, so the screening finds nothing and reports nothing wrong.
The framework generates the missing half of the map. A-LSTM and T-LSTM chemical language models, trained on SMILES representations of known fluorochemicals, sampled over 1.4 million candidate structures. Those were deduplicated by canonical InChIKey, checked with Jensen-Shannon divergence to confirm the property distributions matched real fluorochemistry, and screened against diagnostic fragment rules for structural plausibility. The surviving structures form a suspect library 21.6% larger than legacy repositories, fed into the APP-ID workflow to annotate high-resolution mass spectrometry features.
On spiked validation samples it reached 87% top-1 annotation accuracy. Applied to industrial fluorochemical wastewater it annotated 88 novel PFAS features in influent and 100 in effluent, all missed entirely by database matching.
Feature annotation confirms a structure class, not an exact isomer, so distinguishing isomers still requires physical synthesis and authentic standards. Model output also varies with sampling temperature and SMILES canonicalization choices. The authors address the dual-use tension directly, since a model that generates plausible novel PFAS structures for detection generates plausible novel PFAS structures generally, and they published structural metadata under environmental safety guidelines rather than everything.
A study in the Journal of Chemical Information and Modeling, peer reviewed, built a quantum-mechanical-to-AI surrogate pipeline for halogen-pi interactions, trained on nearly 1.4 million MP2/TZVPP single-point calculations of halobenzene-benzene complexes.
Halogen bonding through sigma-hole interactions is difficult territory. Standard force fields and cheap DFT handle it poorly without expensive dispersion corrections, and the accurate methods are too slow for screening. The team represented non-covalent geometries with lightweight geometric descriptors and trained a feedforward network to map configuration directly to interaction energy, capturing sigma-hole charge distributions and dispersion forces.
The surrogate reached near-CCSD(T) accuracy against quantum mechanical reference calculations while running eight orders of magnitude faster than the baseline MP2 calculations. It held performance on independently generated test sets and on protein structures pulled from the Protein Data Bank, which is the harder test since real protein geometries are not drawn from the training distribution.
Parameterization is narrow by design, covering halogen-pi systems within defined geometric boundaries. Highly conjugated or organometallic halogen systems need supplementary training. Within scope, sub-kJ/mol accuracy at negligible cost makes real-time screening of large libraries possible for kinase inhibitor optimization and organic semiconductor crystal engineering.
The University of Girona published a pedagogical framework in the Journal of Chemical Education, peer reviewed, integrating LLMs with bibliometric mapping into chemical engineering coursework, using ChatGPT, Gemini, CAS SciFinder, and VOSviewer.
The course teaches Boolean prompt engineering, LLM-assisted data extraction, and PRISMA systematic review protocols. Students curated over 50 studies on nanocellulose emulsions, assessed partly through a reflective memo worth 10% of the grade.
What distinguishes this from a general AI literacy course is that it names its target: epistemic drift, meaning the gradual shift in what a field believes caused by uncritical reliance on LLM-extracted summaries and automated meta-analyses. That is a more precise worry than hallucination. A fabricated citation gets caught eventually. A subtly skewed synthesis of real papers propagates into the next literature review and the one after that.
The framework requires active instructor monitoring to prevent reliance on hallucinated citations, which means it does not reduce faculty workload and is not meant to.
DP Technology with collaborators at Princeton, Peking University, and Xiamen University released DPDispatcher in the Journal of Chemical Information and Modeling, peer reviewed. It is an open-source Python framework for fault-tolerant task scheduling across HPC clusters, with local and SSH backends and automatic task retries.
This is plumbing, and it matters because of what now sits on top of it. LLM agents launching large-scale materials screening or active learning potential generation need to dispatch thousands of jobs and survive individual failures without a human restarting anything. It has been adopted by more than 10 major scientific software packages.
Performance depends on the cluster's underlying scheduler configuration, SLURM or PBS, so it inherits whatever those environments impose. An autonomous research agent that cannot reliably queue a job is not autonomous, and this is the layer that had been missing.
ExxonMobil Technology and Engineering Company published a perspective in Industrial & Engineering Chemistry Research, peer reviewed, arguing that first-principles models remain mandatory for energy transition scale-up.
The argument is about extrapolation. Purely empirical models fail when predicting behavior in novel reactor scale-up regimes, because scale-up is by definition a request for a prediction outside the data. The paper supports this with two industrial case studies in fuel cells and carbon capture, showing scale-up risk reduction from retaining physical models.
This is a qualitative perspective backed by case studies rather than a new benchmark, and it comes from a company with commercial interests in the conclusion. It is also consistent with what the rest of 2025 kept finding, from the April combustion work through August's refinery comparison, and it carries weight precisely because it comes from people who own the reactors that fail.
Researchers at Politecnico di Milano, the Italian Institute of Technology, and the University of L'Aquila published an inverse parameter condensation method in the Journal of Chemical Information and Modeling, peer reviewed, that uses machine learning to predict and condense biomolecular force field parameters.
The method captures the statistical variability of force field parameter distributions rather than collapsing them to single values, delivering a 30-fold increase in computational efficiency with minimal loss of structural accuracy. Accuracy degrades slightly on highly atypical non-standard heterocycles, which is where force fields have always been weakest.
November closes the year on the argument 2025 kept making from every direction. The self-driving lab found an optimum that single-variable tuning structurally cannot reach. ExxonMobil said empirical models break exactly where scale-up needs them. A generative model found 188 pollutants that database matching missed, and a curriculum warned that the same generative capability quietly reshapes what a field believes it knows. The tools got dramatically better this year at searching spaces humans cannot. What they still cannot do is tell you when they have wandered outside the one they were shown, which is why every good result this month arrived bolted to something physical.
Questions about this issue or the underlying research
info@cheairesearch.com