April was the month the field stopped adding parameters and started adding rules. MIT modeled all of hydrogen combustion with 344 of them. A 7-billion parameter chemistry model beat the general-purpose giants by knowing one thing well. A quantum computing paper worked out exactly how many logical qubits real chemistry needs, and the number is not one anybody currently has. Most of this month's work arrived as preprints, so read the results as promising rather than settled.
On April 25 researchers introduced Chemma, a 7-billion parameter model adapted from LLaMA-2 for synthetic organic chemistry, trained on 1.28 million reaction instruction pairs. The work is a preprint. It handles forward reaction prediction, single-step retrosynthesis, yield prediction, condition generation, and experimental optimization from one model.
On template-free single-step retrosynthesis with reaction class unknown, Chemma reached 72.2% top-1 accuracy on the USPTO-50k test set, using the standard 40,000 train, 5,000 validation, 5,000 test split. That beats the NAG2G transformer baseline at 55.1% by 17.1 percentage points, and it beats GPT-3.5, GPT-4, and base LLaMA-2-13B under identical conditions.
The wet-lab result is the more interesting one. Chemma's predicted yields set the prior distribution for Bayesian optimization across an open reaction space, and active learning selected which ligand and solvent combinations to physically test. On an unreported Suzuki-Miyaura cross-coupling of cyclic aminoboronates with aryl halides, the loop identified tricyclohexylphosphine and 1,4-dioxane within 15 active-learning iterations and delivered a 67% isolated yield. No quantum-chemical calculations and no handcrafted reaction rules were involved.
The representation is the limit. SMILES strings are one-dimensional and carry no explicit 3D geometry, stereochemical constraint, or transition-state energetics, so the model produces synthetically impractical precursors on dense natural products with multiple chiral centers. The dual-use concern is stated plainly in the source: a generative retrosynthesis model can plan routes to scheduled substances and hazardous precursors, and needs API-level filtering against chemical weapon databases. A model that found a working ligand in 15 tries is a model that would find other things in 15 tries.
MIT researchers Benjamin C. Koenig, Suyong Kim, and Sili Deng introduced ChemKANs in an April 18 preprint. The architecture swaps standard multi-layer perceptrons for Kolmogorov-Arnold Network Ordinary Differential Equations to model stiff chemical kinetics.
The structural difference is where the learning happens. Standard networks put fixed activation functions like ReLU at the nodes and learn linear weights between them. ChemKANs put learnable gridded B-spline basis functions directly on the edges. On top of that the model carries a two-tier inductive bias. A Kinetic Core maps species concentrations to mass fraction production rates. A Thermodynamic Superstructure then enforces mathematical coupling between species generation rates, net enthalpy changes, and energy conservation, with soft constraints on elemental atom conservation. Stiff integration runs through forward sensitivity analysis.
The parameter count is the result worth stating twice. ChemKANs modeled the complete 9-species hydrogen-air combustion system plus temperature in a single network of 344 parameters. The ChemNODE baseline needed 7 separate sub-networks totaling 637 parameters, and it had to drop 3 critical radical species to stay stable. ChemKANs also hit a 2.0 times speedup over the detailed chemistry solver in Arrhenius.jl while keeping full species resolution, and on synthetic biodiesel pyrolysis data carrying up to 15% added noise it showed no overfitting where DeepONet overfit to the noise spikes.
It degrades in low-temperature initiation regimes, where radical generation rates jump exponentially at a threshold and the spline grid is too coarse to follow. That is a real gap for cold ignition modeling. The governance note attached is not decorative: surrogate neural ODEs used in plant digital twins must guarantee bounded outputs, because an unbounded surrogate in a thermal simulation is how a digital twin invents a runaway that the real reactor then has no warning about.
A team presented HPC optimizations for MACE, a 3D graph neural network chemistry foundation model, accepted at ACM HPDC 2025 and posted as an April 14 preprint. The problem was not the model. It was that nobody had written a distributed training setup that understood what molecular data looks like.
Standard distributed training splits a dataset by sample count. Molecules do not cooperate with that, because a 3D molecular graph varies enormously in atom count and interaction density, so equal sample counts produce wildly unequal compute and memory loads across GPUs. The authors reformulated data distribution as an iterative multi-objective bin-packing problem, balancing total atom count and interatomic interaction density across ranks. They also rewrote the high-order equivariant tensor products, the model's main bottleneck, to run symmetric tensor contraction on hardware-native matrix multiplication units.
Across 740 GPUs pretraining on a 2.6 million sample dataset, per-epoch time fell from 12 minutes to 2 minutes, a 6-fold speedup, with near-linear load balance and no memory overflow faults.
The honest framing, which the authors give themselves, is that none of this makes MACE a better physicist. Expressive power and interatomic potential accuracy are unchanged. What changed is who can afford to train one, which the dossier ties directly to reducing compute monopolization by large technology firms. The bottleneck in chemistry foundation models turned out to be a packing problem.
On April 8, Felipe H. da Jornada and co-authors published an end-to-end algorithmic framework, as a preprint, for simulating full-particle chemical dynamics on fault-tolerant quantum computers.
The method attacks the memory scaling problem with pseudoions, which consolidate chemically inert core electrons and atomic nuclei into single dynamical entities, extending classical pseudopotential theory to explicit two-body quantum interactions. The framework builds first-quantized plane-wave representations and block-encodes Hamiltonian interactions using quantum rejection sampling. DFT calculations combine with machine-learning-derived molecular identification traits, coherently encoded into quantum states for amplitude estimation.
What is new is scope. This is the first rigorous fault-tolerant quantum algorithm accounting for coupled quantum degrees of freedom across both electrons and nuclei, which means it moves past the Born-Oppenheimer approximation that treats nuclei as fixed. For a complete 40-particle donor-acceptor reaction, the resource calculation comes out to 808 logical qubits and a specified Toffoli gate count per femtosecond of simulated time.
Those are logical qubits, which require full error correction and gate fidelities orders of magnitude better than current noisy intermediate-scale hardware delivers. The dossier frames the value of this precisely: a clear algorithmic mapping prevents deceptive claims about near-term quantum utility in chemistry. Publishing the exact cost of the thing is the most effective way yet devised to stop people implying they have already done it.
Yuqi Li and collaborators published an arXiv preprint on April 4 presenting a chemistry-aware model that predicts battery degradation under dynamic, non-standardized power profiles.
Conventional battery health models train on clean constant-current, constant-voltage cycles, which is not how any battery in a vehicle or a grid installation is ever used. This architecture couples Hidden Markov processes simulating realistic non-periodic power demand with an automated batch-testing platform. High-throughput X-ray Photoelectron Spectroscopy data mapping Solid Electrolyte Interphase composition feeds an automated polynomial feature extraction algorithm that turns irregular voltage curves into diagnostic chemical indicators.
The model predicted critical capacity knee points and resolved 6 distinct chemical failure modes inside the SEI layer, all inferred from macroscopic electrical signals rather than a destructive teardown.
Calibration is chemistry-specific, so an LFP-tuned model does not transfer to NMC, and solid-state electrolyte interfaces require retraining outright. The safety value is early identification of SEI decomposition knee points, which is the failure pathway that precedes thermal runaway. Reading a battery's internal chemistry from the noise in its voltage curve is a genuinely useful trick, and it is worth remembering the model only knows the chemistries it was calibrated on.
The U.S. National Science Foundation awarded a five-year, $15 million renewal to the Molecule Maker Lab Institute at the University of Illinois Urbana-Champaign. The funding covers foundational AI agents, the AlphaSynthesis platform, modular LLMs, and accessible digital education tools.
The education component is not an afterthought in the award. The MMLI Escape Room and Digital Molecule Maker, both gamified platforms aimed at K-12 and undergraduate AI-chemistry literacy, scale up under the renewal. Whether an escape room teaches retrosynthesis better than a problem set is an empirical question nobody has answered, but the funding structure at least treats it as one worth asking.
Executive Orders 14277 and 14278, issued in April, established federal task forces for AI education and technical workforce retraining, and funded cloud-automated laboratory infrastructure across NSF, DOE, and NIST. These are policy directives, so the outcomes are commitments rather than results.
Published work in the Journal of Chemical Education covered the same ground with more specificity. The AI4CHEM curriculum is a full-semester course for synthetic chemists, built around a zero-install web workflow platform where students build, visualize, and evaluate machine learning models without setting up a local coding environment. Its focus is chemical contextual data representation, which is the part most general machine learning courses skip and the part chemists most need.
Removing the install step is a smaller idea than a federal task force and it addresses a real barrier. A semester spent debugging a Python environment is a semester not spent on chemistry.
April's throughline was that structure beats scale. The winning entries this month were a 344-parameter network carrying thermodynamic laws, a 7-billion parameter model that knew chemistry instead of everything, and a data loader that understood molecules are uneven. The quantum paper fits the same pattern from the other direction, telling everyone precisely what the real problem costs instead of implying it had been solved. Most of these results are preprints and most of them still need someone to check the chemistry in a physical lab. The field spent April learning that the useful move is usually telling the model what it is not allowed to forget.
Questions about this issue or the underlying research
info@cheairesearch.com