February was the month AI stopped answering questions and started holding debates. Google built a system where specialized agents argue over hypotheses in ranked tournaments until one survives. A drug design platform optimized three chemical domains at once and produced a candidate in under half the usual time. Meanwhile the Royal Society of Chemistry retracted a review containing 48 fabricated citations, which is the other thing that happens when a language model is left to generate text unsupervised.
On February 19, Google DeepMind unveiled AI Co-Scientist, a multi-agent system built on Gemini 2.0. It does not run one model. It runs a coalition of specialized agents under a central Supervisor Agent that allocates compute and manages shared memory context.
Each agent has a job. The Generation Agent searches literature and breaks a research objective into foundational testable statements. The Reflection Agent decomposes each hypothesis into sub-assumptions and runs step-by-step checks on novelty, plausibility, and experimental feasibility. The Ranking Agent runs pairwise Elo-based tournaments between competing hypotheses, using simulated scientific debates to assign numeric quality scores. The Evolution Agent takes the tournament winners and generates variations by injecting multidisciplinary perspectives or restructuring assumptions. A Proximity Agent clusters outputs to kill redundant hypotheses and prevent mode collapse, and a Meta-Review Agent reads the system's own trace logs to tune how the agents interact over time.
The underlying mechanism is test-time compute scaling. Instead of relying only on what the model learned in pre-training, the architecture spends more computation during inference, running extended cycles of proposal, critique, and revision before it commits to an answer.
In a benchmark run with researchers at Imperial College London, the system investigated bacterial gene transfer mechanisms tied to antimicrobial resistance. It processed over 28,000 academic studies, formulated 143 distinct hypotheses, and correctly prioritized the mechanism governing phage-inducible chromosomal islands in 72 hours. Human scientists took ten years of laboratory work to establish the same result.
The results come from preprints and an official release rather than a completed peer-review cycle. The system also has limited access to negative results, which is a real gap, because knowing what failed is most of what keeps a scientist from repeating it. It also inherits the hallucination behavior of the language models underneath it. Google's answer to a model that makes things up was to build six more of them and have them cross-examine each other, and the strange part is that it worked.
Insilico Medicine advanced D16-M1P2, a dual-action PROTAC targeting the kinase PKMYT1, to pre-candidate stage. PROTACs are harder to design than ordinary small molecules. A working one needs three things optimized simultaneously: a warhead that binds the target, a moiety that binds an E3 ubiquitin ligase, and a linker of the right length and geometry connecting them. Optimizing any one of those tends to degrade the other two.
The Chemistry42 platform ran the whole chain. It merged pharmacophore features from two known kinase inhibitor classes to generate 2,023 candidate warhead molecules, then applied 40 integrated machine learning models scoring binding affinity, synthetic accessibility, and ADMET profiles to select one for synthesis. It then designed flexible linkers to tether that warhead to the E3 binder at a geometry that forms a working ternary complex.
Selectivity is where the result lands. Profiled against a panel of 403 human kinases, D16-M1P2 inhibited only 4 non-target kinases. Preclinical xenograft models showed tumor regression in CCNE1-amplified tumors, and the molecule showed favorable oral bioavailability across multiple animal species. The program reached pre-candidate validation in 12 to 18 months, against an industry norm of 3 to 5 years, synthesizing fewer than 200 physical molecules where typical projects synthesize 2,000 to 5,000.
The caveats are substantial and worth keeping attached to the number. In vivo clinical efficacy and human safety are unevaluated. PROTACs routinely exceed 800 Da, which brings membrane permeability, metabolic stability, and synthesis scalability problems that only clinical work resolves. Chemistry42 is also a closed commercial platform, so academic labs cannot reproduce the design pipeline, and the dossier flags reporting of training distributions and synthesis routes as a live reproducibility concern.
AutoSID published online in Industrial & Engineering Chemistry Research on February 12, extending the automated system identification work accepted the previous month. The February record adds detail on the search machinery. The pipeline ingests industrial time-series streams, runs preprocessing and dynamic delay alignment, then hands candidates to a structural model search engine deploying four separate hyperparameter optimization algorithms, including Bayesian Optimization and Tree-structured Parzen Estimators, across three selection criteria. The candidate space covers classical state-space and ARX models, tree ensembles including XGBoost and Random Forests, and deep recurrent architectures including LSTMs, GRUs, and Neural ODEs.
The finding held across all 11 process systems engineering case studies. Balanced-complexity tree ensembles beat both simplistic linear models and deep recurrent networks on multi-step predictive accuracy and training efficiency. The specific failure mode is worth naming: deep networks repeatedly overfitted to process noise in low-data regimes, which is exactly the regime most plants operate in.
AutoSID handles offline identification and tuning. Real-time integration with a physical distributed control system needs low-latency infrastructure and fail-safe fallback algorithms that this framework does not provide. It shipped under open-source Creative Commons licensing with full code transparency, which the dossier ties directly to the reproducibility problem in data-driven process modeling. The suggested angle in the source material is to stop over-engineering digital twins, and the benchmark supports it.
Industrial & Engineering Chemistry Research published Control-Informed Reinforcement Learning on February 20. Pure reinforcement learning agents learn process control through trial and error. They are flexible and they are also unpredictable during transient disturbances, which is why plant managers do not let them near a reactor.
CIRL restructures the policy rather than fencing it. The policy output combines a classical PID controller's baseline signal with a deep neural network's residual correction, embedding the PID error, integral, and derivative terms into the inner layers of the policy network's computation graph. This is not reinforcement learning tuning PID constants from outside, and it is not a PID filter cleaning up the agent's output afterward. The baseline behavior conforms to classical control stability principles by construction.
Benchmarked against unconstrained DDPG agents, PPO models, and standalone industrial PID controllers on non-linear chemical reactor tasks, CIRL showed faster setpoint recovery, zero steady-state tracking offset, and reduced control actuator wear. It also reached optimal convergence with up to 60% fewer training episodes.
The framework needs workable baseline PID constants before training starts. For highly coupled multivariable systems where classical loop tuning is already hard, that setup cost grows. The safety case is the real point here. Bounding policy outputs deterministically is what makes these agents compatible with OSHA Process Safety Management standards, and it limits the operational risk of an unconstrained agent exploring a live plant. Evaluation so far is on simulated environments, with hardware-in-the-loop pilot plant testing still ahead.
Research in the Journal of Chemical Information and Modeling showed that machine learning potentials transfer across elements in the same periodic group. Machine learning potentials sit between expensive DFT calculations and cheap classical molecular dynamics. Training a good one takes a large DFT dataset covering many structural phases and thermodynamic states, which is precisely what rare materials do not have.
The workflow trains a deep neural network potential on an extensive base dataset of silicon potential energy surfaces, capturing atomic interaction topologies and both bulk solid and liquid phase behavior. The hidden representation layers holding general group interaction structure are then frozen or used as initialized weights, and the model is fine-tuned on a sparse germanium dataset.
Transfer-learned models beat models trained from scratch on identical sparse data, posting lower force-prediction RMSE across solid and liquid bulk regimes. In small data regimes they also predicted the Phonon Density of States accurately and held simulation stability across temperature sweeps. Models trained from scratch on the same sparse data exploded.
The effect depends on shared electronic configuration. Performance drops sharply transferring across disparate periodic groups, so silicon does not teach a transition metal anything useful. Within Group 14 it works, and it cuts the DFT hours that previously gated computational materials work for smaller research groups.
On February 26 the Royal Society of Chemistry retracted a tutorial review on spectroscopic methods for organic pollutant determination in natural waters, published in Environmental Science: Advances. The retraction notice cites undisclosed generative AI text composition and 48 corrupted or entirely fabricated academic citations.
The sequence matters more than the count. The authors drafted text with an unvetted generative AI tool that fabricated citations. The paper cleared peer review and published. Independent scientific integrity researchers audited it afterward and flagged the reference list to the editors. During the formal inquiry the authors could not produce verification sources or explain where the citations came from, and offered to replace the 48 corrupted entries instead. The Executive Editor retracted it.
The failure that should worry people is the one in the middle. Peer reviewers did not catch 48 fabricated references, because plausible-sounding citations are exactly what a language model is good at producing and exactly what a human reader skims past. Publishers including the RSC, ACS, and Elsevier are now deploying software that cross-checks manuscript bibliographies against Crossref and PubMed before peer review. The month also carried a separate citation manipulation investigation that reached 35 retractions across chemistry journals for systemic self-citation inflation and image duplication. A review about measuring pollutants in water turned out to have a contamination problem of its own.
Chemical educators published a redesigned undergraduate organic chemistry curriculum in the Journal of Chemical Education that pairs commercial generative AI tools with the 12 Principles of Green Chemistry. The setting is a classic microscale cyclohexene synthesis by acid-catalyzed dehydration.
The design decision is the interesting part. Rather than banning chatbots to prevent cheating, the curriculum treats them as raw hypothesis generators whose output students are required to audit. Students write targeted prompts asking ChatGPT, Gemini, and Copilot to suggest greener acid catalysts in place of concentrated sulfuric or phosphoric acid. They then cross-examine those suggestions against environmental impact metrics, health hazard criteria, atom economy, and what the bench equipment can actually do. They run the microscale synthesis with their chosen setup. Finally they audit the original AI output and document its hallucinated claims, missing safety steps, and unrealistic yield expectations.
Pre- and post-course surveys showed improvement in understanding of green chemistry metrics and an increase in students' critical confidence when auditing digital tools for factual reliability.
The structural problem is vendor drift. Commercial models change continuously, so a prompt that produced a useful wrong answer in the fall may produce a different wrong answer in the spring, and instructors have to keep rewriting grading rubrics against a moving target. Grading an assignment whose source material updates itself without warning is a new category of academic labor.
Computers & Chemical Engineering published an MLOps framework for real-time soft sensors in industrial fed-batch penicillin fermentation. Soft sensors predict hard-to-measure biological variables like real-time biomass and penicillin concentration from easy online measurements like pH, dissolved oxygen, temperature, and aeration rate. Static neural network soft sensors decay in service, because batch-to-batch biological variability, raw material lot changes, and sensor fouling all shift the process out from under a fixed model.
An LSTM network ingests the time-series streams and produces continuous concentration predictions. Monitoring algorithms compute residuals between offline lab assay samples and the online sensor output. When those residuals breach pre-set statistical process control bounds, the pipeline pulls recent batch history, preprocesses it, retrains the LSTM weights, validates the new model against safety bounds, and redeploys it. No human is in that loop.
Across 100 simulation batches on the standardized industrial IndPenSim benchmark, the pipeline eliminated model drift and held accurate predictions through multi-month operational cycles. Testing ran on the benchmark simulator rather than an operating physical plant, which is the gap between a self-healing sensor and a self-healing sensor someone has trusted with a real fermenter.
February's throughline was AI systems that check each other, and the month made the case in both directions. Where the checking was designed in, results got better: agents debated hypotheses into shape, a PID controller kept a reinforcement learner inside safe bounds, a soft sensor caught its own drift, and students audited a chatbot instead of copying it. Where nobody checked, 48 citations to nonexistent papers went through peer review and into the permanent record. The tooling to argue with a language model existed all month. The retracted review just did not use any of it.
Questions about this issue or the underlying research
info@cheairesearch.com