January was the month AI science stopped pretending the model was the whole system. Governments funded shared laboratories, engineers put physics back into process models, and chemists put language-model reasoning inside optimization loops. A chemistry exam then found that ChatGPT handled the word "NOT" better than the students. This is progress, with a very specific definition of progress.
In January, the US Department of Energy, Japan's Ministry of Education, Culture, Sports, Science and Technology, and the Ministry of Economy, Trade and Industry announced a $1 billion bilateral partnership under the US Genesis Mission. Each country will commit $500 million over five years. The plan joins twelve US National Laboratories with twelve Japanese research institutions, including RIKEN, NIMS, and the University of Tokyo.
The technical plan is an autonomous research network. AI agents will run across high-performance computing systems such as DOE user facilities and Fugaku. Standardized API layers will connect them to robotic labs, high-throughput characterization equipment, and neutron or synchrotron beamlines. Streaming results will feed active-learning foundation models that select the next physical trials. This is not facility sharing. It is a shared loop.
The partnership includes eleven joint scientific projects. They cover autonomous material synthesis for clean energy and quantum devices, particle accelerator controls, and semiconductor material discovery. The baseline is slower and more human. Traditional collaboration uses asynchronous proposal cycles, manual data synthesis, and local lab operations. The Genesis Mission targets a 2x increase in overall scientific productivity within five years.
That ambition has conditions. The funding remains subject to annual appropriations from the US Congress and the Japanese Diet. The technical protocols must balance open collaboration with export restrictions on advanced AI model weights and dual-use chemical or materials data. Cabinet-level governance includes safety verification layers, and human sign-offs remain required before automated synthesis of bio-hazardous or high-risk explosive chemicals. The agents can plan the work. Humans still approve the dangerous part.
For industry, the model points toward materials discovery, carbon capture catalyst synthesis, and advanced nuclear fuel modeling tied to high-performance computing. The next test is late-2026 benchmark deployments of robotic synthesis pipelines between national labs, along with operating guidelines for shared supercomputing allocations on Fugaku and Frontier. The robots have a schedule now. The appropriations still arrive one year at a time.
Adrian Caspari and colleagues published a peer-reviewed study in Computers & Chemical Engineering on incremental identification of dynamic hybrid models. The problem is familiar. If engineers train a first-principles process model and a machine-learning component at the same time, parameter estimation can become slow, unstable, and hard to interpret.
The paper separates the work into a four-stage pipeline. Regularized dynamic parameter estimation first finds time profiles for unmodeled flux variables. Correlation analysis then links those variables to internal process states. Separate machine-learning models fit the residual dynamics. The system finally recouples the machine-learning surrogate with first-principles mass and energy balances. The machine gets the messy part. Mass and energy still get the final word.
The method was tested on three complex industrial dynamic process case studies. It cut optimization solving times by up to an order of magnitude compared with simultaneous nonlinear programming solvers. It maintained predictive dynamic accuracy under sparse time-series sampling and noisy sensor conditions. The comparisons included traditional simultaneous hybrid identification and pure black-box dynamic neural network architectures.
That matters for digital twins in chemical refining, bioprocess batch reactors, and polymer manufacturing. Engineers know the conservation laws, but not every kinetic, fouling, or microbial-growth detail. The physical component also blocks unphysical predictions during wild operational transients. Deployment still needs a high-quality initial physical model, dynamic data, and domain expertise for the correlation analysis. The next watch point is integration into Aspen Plus and gPROMS for real-time model predictive control. AI has been allowed into the plant. It has not been allowed to repeal mass balance.
On January 12, a ChemRxiv preprint from Sun Yat-sen University, Xiamen University, and Guangzhou Laboratory introduced AIQM-PBSA. It combines AIQM3 machine-learning interatomic potentials with Poisson-Boltzmann surface area calculations for biomolecular binding free energies. It is a preprint, so the result has not cleared peer review.
AIQM-PBSA uses an ONIOM partitioning scheme. AIQM3 handles the ligand and active-site residues in the core binding region. Classical molecular mechanics handles the surrounding protein and solvent. PBSA supplies implicit-solvent free-energy terms, and interaction entropy accounts for conformational flexibility. Traditional MMPBSA uses classical empirical force fields, which can struggle with charge transfer, metal coordination, and severe electronic polarization. AIQM-PBSA replaces the MM intramolecular energy term with a machine-learned quantum potential without the scaling cost of full quantum-mechanical calculations.
The benchmarks produced real numbers. On CASF-2016 Dataset B, AIQM-PBSA achieved up to a 31% higher Pearson correlation with experimental binding affinities than standard empirical MMPBSA. It achieved a 16% higher Pearson correlation than ANI-2x-based binding calculations. On the Schrödinger JACS benchmark, it evaluated 87 complexes across CDK2, JNK1, P38, and TYK2. It distinguished activity variations among congeneric ligand series.
That gives drug discovery and enzyme design a better way to rank lead compounds before synthesis. The preprint also has a finite guest list. Current tests cover neutral organic molecules built from H, C, N, O, F, S, and Cl. Charged species and transition-metal complexes need more training. The method is faster than full quantum mechanics but heavier than classical molecular mechanics and needs GPU pipeline orchestration. Dual-use screening for regulated toxins or restricted biological agents also needs controls. Next up are metalloenzymes and membrane-bound receptor environments.
The University of Liverpool's Cooper Group introduced BORA in a ChemRxiv preprint on January 27. BORA means Bayesian Optimization with Reasoning Agents. It couples large language model scientific reasoning with classical Bayesian Optimization to control closed-loop physical and computational experiments. Related January work also used agentic LLM systems for transition-state searches in theoretical chemistry. The optimizer now has a chemistry opinion. It also has an API bill.
Standard Gaussian Process Bayesian Optimization is good at numerical search, but it does not easily use qualitative chemical intuition. BORA sends physical or computational outcomes to an LLM reasoning engine. The engine evaluates hypotheses and guides a hybrid surrogate that combines Gaussian Process Bayesian Optimization with contextual acquisition functions. The resulting plan narrows the search to chemically viable parameters before commands go to experimental hardware.
Against five modern closed-loop strategies, including Gaussian Process Bayesian Optimization, unconstrained LLM optimization, and random search, BORA converged faster toward optimal synthesis conditions. The advantage was strongest in high-dimensional spaces, where solvent compatibility and functional-group stability can remove wasteful trials. The evaluation also reported a higher parameter exploration rate. This is the useful part of giving the optimizer a chemistry filter.
The fine print is practical. LLM output can fluctuate, so schema enforcement and logit constraints must block invalid parameter syntax from reaching robotic dosing units. Frequent calls to a large reasoning model add API cost and latency compared with lightweight mathematical Bayesian Optimization. BORA places hard programmatic safety bounds outside the LLM context window. The next watch point is integration with SiLA 2 and AnIML, plus industrial active pharmaceutical ingredient route optimization. The chemist's intuition is in the loop. So is the invoice.
A January 7 study in the Journal of Chemical Education from Florida State University used binomial logistic regression to compare ChatGPT with human undergraduates on General Chemistry II multiple-choice exams. It analyzed individual item variables instead of reporting one overall accuracy score. The variables included word count, explicit negative phrasing such as NOT, fill-in-the-blank structures, visual prompts, and Bloom's revised taxonomy categories such as interpretation, classification, and algorithmic calculation.
That produced a useful mismatch. Every doubling of prompt word count lowered the odds of a correct answer by 10% to 20% for both students and ChatGPT. On questions with explicit negative constraints, ChatGPT significantly outperformed human students. Humans did better on interpretation and classification items that required multi-step chemical relationships. The model noticed NOT. The students noticed the chemistry.
That does not make ChatGPT a chemistry tutor. Model updates can change its baseline accuracy, and the study covered multiple-choice General Chemistry II items. It did not measure open-ended, handwritten synthesis derivations or physical laboratory skills. The result still gives faculty a design tool. They can use interpretation and classification tasks to make diagnostic exams harder to outsource, or use AI to expose student blind spots.
The study also supports prompt literacy and critical validation. General-purpose LLMs can produce plausible answers while missing complex conceptual interpretation. The next test is broader item-level regression in organic chemistry, thermodynamics, and fluid mechanics across multiple university campuses. The model got the word NOT. It is still working on what the question means.
Questions about this issue or the underlying research
info@cheairesearch.com