June was the month the tooling got given away. Microsoft released a neural functional and the dataset behind it. MIT CSAIL and Recursion open-sourced a binding affinity predictor under the MIT license. Somebody published 3.5 million DFT relaxation trajectories. Also released, though not deliberately, was a measurement of how much fabricated literature is now circulating in the scientific record. Most of the month's work is preprints, so treat the numbers as reported rather than settled.
Microsoft Research AI for Science announced Skala, a neural network-based exchange-correlation functional for Density Functional Theory, alongside the Microsoft Research Accurate Chemistry Collection. The work is a preprint with an official release.
DFT has a standing tradeoff. Semi-local functionals are fast and approximate. Hybrid and double-hybrid functionals are accurate and computationally brutal. Skala goes after the middle by abandoning the hand-crafted localized density features that Generalized Gradient Approximation and meta-GGA functionals depend on. Instead it evaluates electron density, its gradient, its Laplacian, and kinetic energy density across an irregular integration grid, using a neural architecture with grid-point communication mechanisms. That communication is what captures non-local electronic interactions without paying the scaling cost that hybrid functionals demand.
The training set is the other half of the release. MSR-ACC contains 76,879 total atomization energies computed with the W1-F12 thermochemical protocol, two orders of magnitude larger than previous datasets at comparable accuracy, sampling chemical space across all elements up to argon.
Skala reached mean absolute errors below chemical accuracy for small-molecule atomization energies, stayed competitive with top hybrid functionals on the LMGB35, HMGB11, and CCse21 benchmarks, and ran over 100 times faster than DM21 in PySCF. Baselines spanned Jacob's Ladder, including revPBE, r2SCAN, B3LYP, and M06-2X, plus W4-17 wavefunction reference data.
It does not converge cleanly on hard multi-reference systems. Getting SCF convergence in PySCF required a multi-stage retry protocol with DIIS adjustments, level shifting, and damping, and spin-symmetry breaking has to be triggered explicitly on multi-referential structures to get binding energies right. The dataset is open access and the weights were prepared for public release, which is the part that determines whether anyone outside Microsoft benefits from it.
Researchers released ChemGraph, an open-source multi-agent framework for computational chemistry workflows, as a preprint. It is built on LangGraph following the ReAct execution model, and it wraps RDKit, the Atomic Simulation Environment, and PubChemPy as structured agent tools, with an internal AtomsData class serializing molecular geometries across agent state graphs. Through ASE it dispatches to cheap ML potentials like MACE and Meta's UMA, or to NWChem and Orca for high-accuracy quantum mechanics.
The work splits across four roles. A Planner Agent decomposes the user's objective into discrete stages. A Loop Controller routes each command to an Executor Agent, which invokes the actual backends for energies, gradients, or properties. An Aggregator Agent then synthesizes the raw results into a final response.
Across 13 benchmark experiments and 260 independent execution runs, the split mattered most where the task was longest. On simple tasks, small models did fine unaided. On complex reaction thermochemistry requiring 9 to 12 sequential tool invocations, multi-agent decomposition raised task accuracy by up to 73% in GPT-4o-mini and Claude-3.5-haiku, matching or beating single-agent GPT-4o. The mechanism is unglamorous and real: breaking the job across agents prevents the context-window degradation that kills a single agent halfway through a twelve-step calculation.
No model reached 100% accuracy on the complex reaction benchmarks. The framework also inherits its tools' failures, so a DFT or MLIP calculation that will not converge can stall the agent loop entirely without explicit fallback logic. Automated tool execution needs sandboxing too, because an agent framework that runs arbitrary code on a compute cluster is a security question wearing a chemistry hat.
MIT CSAIL and Recursion released Boltz-2 under the MIT license, an open-source model predicting biomolecular 3D complex structures and small-molecule binding affinities at the same time. This is an official announcement rather than a peer-reviewed paper.
Existing tools split the job badly. Structure predictors give you a shape and then you run post-hoc docking. Accurate binding thermodynamics means Free Energy Perturbation, which is physically rigorous and computationally punishing. Boltz-2 extends the AlphaFold 3 approach by integrating structure prediction loss directly with binding affinity regression, modeling the co-folding of protein-ligand complexes and predicting affinity from the co-folded representation rather than from a separate docking step.
It runs on the order of 1,000 times faster than physics-based FEP, benchmarked against FEP workflows, AutoDock Vina, and AlphaFold 2 and 3. Recursion confirmed it is running inside their active internal drug discovery pipeline.
Deep learning affinity predictions still produce false positives on novel binding pockets poorly represented in structural training data, which is exactly where a drug discovery program most wants a trustworthy answer. The MIT license does something specific about that: open weights let the community inspect the model and evaluate biological security risks around target design, which no proprietary structure predictor permits.
ether0 returns this month in a larger form, released as a preprint. It is now a 24-billion parameter model derived from Mistral-Small-24B through post-training reinforcement learning on 640,730 chemistry problems spanning 375 distinct tasks.
The claim is about method rather than scale. There is no domain-specific pretraining on a massive chemical corpus. The reinforcement learning is grounded in physical and chemical ground truths across tasks including synthesizability prediction, blood-brain barrier permeability, human receptor activity, and molecular odor classification. The model emits a natural language reasoning chain before producing structures in SMILES and SELFIES notation.
Evaluated against GPT-4o, Claude 3.5 Sonnet, domain models including ChemCrow and LLaMA-3-Chem, and human domain experts, ether0 surpassed the frontier generalist models and the specialized baselines on molecular design, with better data efficiency during post-training.
The costs carry over from the earlier version. Long chain-of-thought generation raises inference latency and token cost, and output SMILES need strict RDKit validation to catch rare syntax errors. Post-training a general model on verified chemistry problems keeps outperforming pretraining a chemistry model on text about chemistry.
Researchers published PubChemQCR, an open dataset of 3.5 million molecular relaxation trajectories derived from PubChemQC, containing over 300 million distinct conformations labeled with DFT total energies and Cartesian force vectors. It is a dataset release with a preprint.
The gap it fills is specific. Machine Learning Interatomic Potentials need training data covering non-equilibrium geometries to stay numerically stable during molecular dynamics, and most datasets are collections of static ground-state structures. PubChemQCR was built by extracting the raw geometry optimization trajectories themselves, capturing each step from high-energy starting structures down to local minima. The developers benchmarked nine representative MLIP architectures, including graph neural networks and equivariant potentials, on energy and force prediction across the dataset splits.
Training a potential only on molecules already at rest and then asking it to simulate motion was always an odd arrangement. This dataset is the correction, and its value depends on how many MLIP groups actually retrain against it.
Chemical engineers built surrogate models for low-salinity water flooding in sandstone reservoirs, published in a peer-reviewed paper. They generated 10,240 training samples from an in-house numerical simulator modeling multi-ion exchange dynamics, clay swelling, and wettability alteration across a 400-day injection timeline, then trained eight architectures including Random Forest, Decision Trees, Neural Networks, and Support Vector Machines on varied injection water ionic concentrations.
Random Forest posted the best predictive performance, with Neural Networks and Decision Trees close behind. The operational result is that simulations taking hours now run in sub-second time while holding numerical accuracy, which turns enhanced oil recovery optimization from a batch job into an interactive one.
The models are bounded by the ionic concentration ranges and sandstone permeability distributions in those original 10,240 samples. Carbonate reservoirs and unmodeled mineralogies require retraining from scratch. A surrogate is a compression of the simulations it was shown, and it has no way to signal when it has been asked about a reservoir nobody simulated.
The Journal of Chemical Education published a peer-reviewed study of a redesigned organic chemistry lab at CUNY Guttman Community College, integrating ChatGPT, Gemini, and Copilot with the 12 Principles of Green Chemistry for cyclohexene synthesis.
Students worked through a structured 5S prompt framework, standing for Specify, Simplify, Structure, Synthesize, and Summarize, querying chatbots for alternatives to corrosive liquid acids and toxic chaser solvents in cyclohexanol elimination. The critical step is what happens next. Every AI-suggested alternative gets routed to the ChemForward hazard database for toxicity and environmental evaluation before anyone touches it. Only cleared suggestions get run on microscale kits, with yields measured by gas chromatography.
Nine students across three lab groups designed and executed alternative pathways. One group replaced corrosive acid with Amberlyst-15 solid acid catalyst and demonstrated both its reusability and product yield by Vernier gas chromatography. Post-activity surveys showed statistically significant gains in green chemistry awareness, prompt engineering literacy, and critical verification skills.
The cohort is nine students at a two-year institution with limited lab infrastructure, which is a small sample and the study says so. The AI also suggested methods the lab could not perform, including microwave heating the equipment did not support. The governance line here is stated without hedging in the source and deserves repeating: AI recommendations must never be executed in a physical laboratory without prior verification against certified hazard databases. The database check is not a pedagogical flourish. It is the thing standing between a chatbot suggestion and a student handling it.
A bibliometric study in Global Knowledge, Memory and Communication analyzed 3,456 retracted AI and machine learning publications, alongside integrity audits from Retraction Watch, Pangram Labs, and independent researchers.
The retraction picture: 60.5% of AI/ML retractions came from data errors or methodology concerns, and 23.2% involved active falsification. The average detection lag for AI-generated-content-related retractions was 2.77 years, against 1.91 years for traditional misconduct, meaning this category of bad paper stays in the literature roughly nine months longer before anyone catches it.
The contamination numbers are worse. An audit of GPT-4o-generated literature syntheses found 19.9% of generated citations were fabricated outright, and 54.6% of the real citations carried bibliographic errors or invalid DOIs. Pangram Labs ran LLM detection classifiers across 19,490 manuscripts and 75,800 peer reviews submitted to ICLR. Retraction Watch confirmed that over 400 US patents cite retracted scientific papers directly.
Bibliometric work depends on Scopus and Web of Science indexing metadata, which undercounts papers with unflagged AI text or unretracted fraudulent data, so these figures are floors rather than ceilings. In a month when four major research groups gave away their models and datasets for the sake of reproducibility, one in five citations from an unsupervised literature review pointed at a paper that was never written.
June's throughline was openness as infrastructure. A functional, a dataset, a binding predictor, a reasoning model, and 3.5 million trajectories all went public in a single month, and each one lowers the cost of checking somebody else's work. Then the bibliometric audit arrived and put a number on what happens without that checking: 19.9% fabricated citations, a 2.77-year detection lag, and 400 patents built on retracted papers. The tools for verification had a very good month. The verification itself is still optional, which is the part that keeps showing up in the retraction notices.
Questions about this issue or the underlying research
info@cheairesearch.com