May produced the first peer-reviewed head-to-head between language models and expert chemists, and the models won on points. They also learned nothing about when to hesitate. The same month, reinforcement learning cut a sugar plant's energy use by 12.0%, and a force field replaced approximations chemists had lived with for 50 years. The throughline is that physics-constrained systems performed, and the unconstrained one was certain about spectra it had invented.
On May 20, a team led by Dr. Kevin M. Jablonka at Friedrich Schiller University Jena published ChemBench in Nature Chemistry, a peer-reviewed benchmark comparing foundation models against human experts across more than 2,700 chemistry tasks. The human control group was 19 experienced chemists.
The setup was deliberately unkind to the models. They were tested on raw parametric memory with no retrieval augmentation and no external tools, using text completion with annotated tags like START_SMILES to isolate molecular representations. Coverage spanned organic, inorganic, physical, and analytical chemistry, from nomenclature to multi-step reasoning. The human experts answered the identical question set through a web application at chembench.org, with permission to use search engines and computational software on designated problems.
OpenAI's o1-preview scored nearly double the top-performing human expert's overall accuracy across the corpus. Llama-3.1-405B-Instruct came in close to proprietary state-of-the-art models. Performance dropped sharply on knowledge-intensive questions where the models had no retrieval system to lean on.
The finding that matters is not the score. On tasks requiring structural interpretation, predicting NMR spectra or identifying stereocenters, the models produced incorrect answers at absolute probability metrics. They showed zero self-calibration about their own error rates. The human chemists expressed hesitation on ambiguous problems, which is the behavior that keeps a wrong answer from becoming a wasted synthesis.
The governance consequence follows directly. Autonomous synthesis engines cannot take LLM completions straight to hardware without a deterministic verification layer checking structural outputs independently. A model that is right twice as often as an expert and never uncertain is more dangerous in an automated lab than one that is right half as often and says so.
Computers & Chemical Engineering published a peer-reviewed digital twin framework applying Proximal Policy Optimization across four connected units of a sugar manufacturing plant: cane maceration, juice clarification, vacuum evaporation, and sugar crystallization.
The construction matters. The team built First-Principles Mathematical Models for each unit, then linked them through state estimation and sensor fusion into one synchronized digital twin. A model-free PPO agent trained on that twin, managing steam flow rates, cooling water circulation, and cane feed modifiers under fluctuating feed composition. Conventional plants run isolated PID loops per unit, which handle neither multi-variable non-linear dynamics nor the coupling between sequential units.
Against PID baselines, the agent delivered a 12.0% reduction in specific energy consumption, a 3.1% increase in total sugar yield, a 3.6% rise in pure water recovery, and a 57.1% drop in process operational instability.
Safety constraints were embedded in the reward function itself, so the policy cannot exceed maximum pressure or temperature bounds to chase yield. That is the right place to put them, because a reward function that permits a violation will eventually find it.
Validation ran on high-fidelity dynamic simulation, not a physical plant. Real rollouts have to survive sensor drift, heat exchanger fouling, and mechanical interruptions that a simulation does not produce unprompted. The appeal for heavy industry is that double-digit energy reduction arrived without a capital-intensive equipment overhaul, which is a rare sentence in decarbonization.
On May 28, researchers from the University of Cambridge, the University of Oxford, and partner institutions published MACE-OFF in the Journal of the American Chemical Society, a peer-reviewed series of short-range transferable machine learning force fields for organic molecules.
Molecular dynamics has run for over 50 years on classical empirical force fields using fixed point charges and harmonic bond potentials. Those approximations are fast and they fail on complex conformational changes and non-standard organic states. MACE-OFF uses higher-order equivariant message-passing neural networks built on the MACE architecture, trained on high-level quantum mechanical reference data covering conformational space, non-covalent interactions, and diverse organic chemistry. Restricting predictions to short-range atomic environments is what buys the transferability to molecules the model has not seen.
It produced smooth, rapidly converging dihedral torsion profiles for previously unobserved organic structures, and predicted gas-phase conformational energies, crystal lattice dynamics, and condensed-phase liquid densities matching experiment. Baselines were the classical GAFF and OPLS force fields plus high-level DFT reference calculations.
Short range is also the constraint. Long-range electrostatics and macromolecular assemblies require pairing the potential with explicit long-range dispersion corrections. The practical payoff is that nobody has to re-parameterize a force field for each new drug candidate, which removes a step that has quietly gated early-stage virtual screening for decades.
Chemical Engineering Journal Advances published a comparative study from the University of Sheffield, University of York, and University of Southampton evaluating parallel versus series hybridization for continuous flow reactor digital twins. Hybrid modeling gets discussed constantly in chemical engineering, and systematic empirical comparisons between the strategies were scarce.
Two architectures went head to head. Physics-Informed Neural Networks embed chemical kinetic differential equations directly into the network loss function, running in forward and inverse modes. The series alternative combines Smooth Particle Hydrodynamics with a Genetic Algorithm, simulating fluid particle dynamics and optimizing system-level parameters across spatial coordinates.
Each won a different contest. PINNs estimated reaction rate constants with errors under 2% under severe data scarcity, at lower computational cost. The SPH+GA framework reached a coefficient of determination up to 0.91 on spatial hydrodynamic resolution, modeling particle-level fluid mixing.
The tradeoffs are explicit. PINNs need an accurate initial mathematical formulation of the kinetics, so they are only as good as the equations someone wrote down. SPH+GA carries substantially higher computational overhead. The useful output here is a selection rule rather than a winner: PINNs for rapid kinetic estimation in microfluidic rigs, SPH+GA for reactor scaling where fluid mixing dominates. Accurate flow reactor twins also predict thermal hotspots and mass transfer bottlenecks before physical scale-up, which is the difference between a modeling exercise and a safety control for volatile compounds.
On May 8, the Institution of Chemical Engineers Computer Aided Process Engineering Special Interest Group hosted a keynote from Prof. Panagiotis D. Christofides at UCLA presenting a cyber-resilient control architecture. This is an official announcement rather than a peer-reviewed publication.
The framework puts recurrent neural networks and ensemble learning inside non-linear Model Predictive Control, in two tiers. The first tier monitors process sensor streams continuously to detect cyber-attacks or equipment anomalies. The second adjusts MPC constraints to hold stable operation while an attack is underway.
The design premise is the departure. Conventional practice separates operational safety from cybersecurity and treats network security as an IT problem happening somewhere else. Embedding attack detection inside the predictive control loop lets the controller maintain physical stability while sensor feeds are actively tampered with. On pilot-scale chemical deposition processes and additive manufacturing testbeds, the architecture identified false-data injection attacks in real time and prevented thermal runaways and shutdown events.
The detector trains offline on simulated attack vectors, so novel patterns outside that training distribution may pass through undetected. This is the standard limitation of learned anomaly detection and it deserves stating plainly here, because the failure mode is a controller confidently stabilizing a process against readings that an undetected attack is still feeding it.
The Journal of Chemical Education published a study on May 13 evaluating undergraduate chemistry students co-creating pharmaceutical analysis reports with generative AI. The assignment structure treats the model as an imperfect junior drafter whose work requires auditing.
Students generated an initial LLM draft analyzing a pharmaceutical product. They then ran a literature audit against primary peer-reviewed sources, critiqued the AI text, corrected technical errors, added physicochemical justifications, and rewrote conversational AI bullet points into structured technical prose. The comparison baseline was traditional literature reporting assignments completed without AI.
The majority of students identified factual inaccuracies in the initial draft, incorporated physicochemical arguments, and produced professional technical writing from the model's output.
A fraction of students still over-relied on unverified AI text, which the study reports rather than smoothing over. The design principle is sound regardless: an unenforceable ban produces undisclosed use, and an assignment that requires finding the model's errors produces students who have practiced finding them.
The NSF Center for Computer Assisted Synthesis held its annual meeting at Carnegie Mellon University on May 15, showcasing LLM-integrated automated laboratories built with Enamine and using AIMNet2 potentials. This is a departmental announcement rather than a published result.
The stated target is a drop in targeted discovery cost from $10 million to under $100,000. That figure is a projection from a consortium meeting, not a measured outcome, and it should be read as an objective rather than an achievement. Physical synthesis throughput remains constrained by hardware, which is the bottleneck that a language model integration does not remove.
May's lesson is that the constraint is the feature. Every system that performed this month carried physics inside it, whether that was conservation laws in a loss function, safety bounds in a reward function, or quantum reference data in a force field. The one system evaluated without constraints scored highest of all and hallucinated NMR spectra at full confidence. The chemists it outscored knew when to hesitate, which turns out to be the part of expertise that nobody has worked out how to benchmark, and the part an automated lab most needs.
Questions about this issue or the underlying research
info@cheairesearch.com