Incoming Member of Technical Staff at Medra, January 2027 · Ph.D., UCLA, 2026
I build symmetry-aware neural networks for chemistry — E(3)-equivariant models over atoms and trajectories, steerable encoders over molecule images, language models that read 3D directions — along with the multi-scale simulations that feed them and the GPU work that makes them trainable.
Chemistry reaches a model as papers, drawings and simulated atoms. Each gets an encoder that respects its symmetry: none for text, planar rotations for drawings, E(3) for atoms. Alignment puts them in one space. Post-training and a harness turn that space into a co-scientist whose proposals go back to simulation as jobs. GPU work sits under every stage.
Vision: a lab agent that understands chemistry
The same systems as parts of one agent. Hover or tap a part to find it in both figures.
Scroll the diagram sideways →
Today's self-driving labs mostly train an athlete to drill the same moves. A lab agent that understands the chemistry behind each protocol step could plan better and find strategies people don't expect.
Planned: a recursive self-improvement loop
One lit path through a quiet field: the agent improves itself only through steps it can check. Hover or tap to trace it.
Scroll the diagram sideways →
Planned. The agent writes new data, keeps only steps that pass the Lean checker, SQL and protocol checks and simulation, trains on what it kept, and widens the trusted regime only where bench results confirm it — then repeats. A human signs off each round.
I work on machine learning for chemistry: equivariant models, chemistry vision and language models, and agents that run simulations.
VERDICT
Four different structure recognizers vote. When they disagree, the structure is flagged as wrong, with no labels needed. AUROC 0.916. Chemists rated 400 released labels at 0.995 precision. arXiv:2608.22183
Image-to-SMILES post-training
LoRA SFT on three open VLMs reached 59.8% exact match on 331 journal figures, against 58.3% for MolScribe. DPO pushed preference accuracy to 90–97% while exact match fell to 0–2%, so I kept SFT.
Electrolyte co-scientist
Qwen3.6-27B with LoRA on 254,527 task, evidence and tool-use rows. Each MD job is checked before it is submitted, and repeat jobs are caught.
A CN steerable CNN as the image encoder for structure recognition. It handles rotations but not reflections, because a mirror image flips stereocenters.
Text tokens attend to l ≥ 1 irreps through Clebsch–Gordan products. Equivariance error is below 10−9 in fp64.
Agents for science
Failure-gated force-field harness
Recorded failure modes of past polarizable force-field fits become automated gates; a candidate field that repeats a known short-range collapse or typing break is rejected before use.
Consensus labels for self-training
Independent structure readers vote on literature figures; only agreed labels enter training, and real crops, not more synthetic drawings, moved accuracy.
Multimodal molecular retrieval
A molecular knowledge graph keyed on InChIKey, with identity, structure-similarity and text search linking literature, structures and simulations.
Bounded scientist agents
Tool-using electrolyte and cell scientist agents that inspect trajectories, diagnose failures and prepare jobs; submission is default-deny and model text never runs as a shell command.
Cell-scale simulation
Large P2D and 3D electrothermal cell simulation campaigns for design studies.
News
Nov 19, 2026
Excited to share that I will join Medra as a Member of Technical Staff in January 2027.
Post-training and evals for chemistry vision and language models, multi-scale electrolyte and cell simulation, and co-scientist agents that plan and run jobs. Notes: Overview · Molecules · Electrolytes · Cells.
2022 — present
UCLA Ph.D. Research
DFT and multi-scale modeling of electrochemical interfaces, electrode degradation, and reaction selectivity; LLM- and vision-based tooling for computational chemistry.
Production ML pipelines for neural network potentials and scalable scientific computing workflows; developer-community and documentation work in the DeePMD ecosystem.
Multi-scale simulation (DFT–kMC), machine learning for catalysis, and electrochemical experiments on bifunctional catalysts for Li–S and Zn–air batteries; three first-author publications.
Selected publications
OCSR VLM
Real Data Closes Synthetic-to-Real Gap in Optical Chemical Structure Recognition
Yani Guan, Dengpan Dong, Zi Wei, and 4 more authors
arXiv preprint arXiv:2608.09100, 2026
* Equal contribution. Vision-language models for image-to-SMILES: real labeled depictions, not more synthetic data, are what close the synthetic-to-real gap.
Millions of chemical structures appear in patents and papers only as drawings. Optical chemical structure recognition (OCSR) appears nearly solved on synthetic images, but remains difficult on real documents: the starting recognizer Qwen2.5-VL-7B reaches over 91% accuracy on synthetic images and below 16% on the ACS, CLEF-IP and USPTO benchmarks. We fine-tuned 21 recognizers on mixtures of synthetically rendered structures and labeled real depictions from patents, journal figures and hand-drawn collections, varying the VLM base, the fraction of real training data, and the vision-tower adaptation strategy. Adding labeled real training images makes the largest difference: for Qwen2.5-VL, ACS exact match rises from 0.15 with no real data to 0.37 at 9.5% real data and 0.46 at 50.2%. A vision LoRA does not help Qwen at all (+0.00, paired p = 1.00), substantially helps InternVL3-8B (+22.8 to +34.6 pt) and modestly helps GLM-4.1V-9B, so its value depends on the base model. The strongest model reaches 0.96 exact match on clean renders and 0.49, 0.65, 0.84 and 0.76 on ACS, CLEF-IP, UOB and USPTO. Differences between bases are largest without real data (0.21) and shrink at 70% real data (0.06), while their ranking also changes – base model and real-data mixture must therefore be selected together.
VERDICT
VERDICT: Agreement Beats Pixel-Space Verification in Real-Document OCSR
Yani Guan, Dengpan Dong, Shuang Luo, and 6 more authors
arXiv preprint arXiv:2608.22183, 2026
Label-free verification for image-to-SMILES at scale: agreement among four architecturally distinct recognizers reaches AUROC 0.916, while re-rendering a prediction and comparing pixels barely beats chance at 0.547.
Optical Chemical Structure Recognition (OCSR) converts 2D molecular depictions in the published literature into SMILES, and is increasingly important for constructing large-scale chemical training datasets. Automation at that scale requires identifying unreliable predictions in the absence of ground truth. Three families of label-free signals were compared on 263 ACS journal depictions with verified ground truth: model confidence, re-rendering similarity, and agreement among recognizers. Pixel-space re-rendering performed little better than chance (AUROC 0.547, 95% CI [0.465,0.629]), and an oracle-tuned threshold on it reduced correct labels per image from 0.745 to 0.205. Agreement among four architecturally distinct recognizers instead reached an AUROC of 0.916 ([0.880,0.952]). The two-of-four rule accepted 81.7% of images at 88.8% precision, the three-of-four rule 52.1% at 98.5%. The same pattern held on CLEF-IP, UOB, and USPTO. This distinction is obscured on synthetic benchmarks, where re-rendered predictions naturally resemble their inputs. A substance filter removed 2,193 false agreements on wildcards and R-group fragments, after which the three-of-four rule rejected all 68 generic depictions. VERDICT was then applied to PMC Open Access, producing 6,146 structure labels for 4,833 molecules; chemist adjudication of 400 released labels in two independent samples yielded precisions of 0.995 for the three-of-four tier and 0.958 for the two-of-four tier. VERDICT therefore enables validated labels for multimodal molecular databases linking structure images, machine-readable representations, and source-publication information. In SES AI’s Molecular Universe platform, VERDICT further serves as an image-based interface for searching and retrieving molecular records.
MU-PFF MD
High-throughput Molecular Dynamics Simulation on an AI-powered Platform
Dengpan Dong, Yani Guan, Shuang Luo, and 5 more authors
ChemRxiv preprint, 2026
MU-PFF: a machine-learned many-body polarizable force field plus automated system building and trajectory analysis, taking an electrolyte formulation from SMILES to predicted properties without manual parameterization.
We present an end-to-end, AI-powered platform for high-throughput molecular dynamics (MD) simulation of liquid electrolytes and complex formulations. At its core is MU-PFF, a many-body polarizable force field whose functional form follows the well-validated exp-6 polarizable model, but whose parameters are produced entirely by an automated machine-learning pipeline rather than by hand. We assemble a quantum-chemistry dataset of roughly one million molecules and molecular complexes and train a single message-passing neural network to predict the complete parameter set – partial charges, atomic polarizabilities, bonded (bond, bend, torsion) terms, and the repulsion-dispersion parameters – directly from molecular structure. Parameterization that traditionally takes weeks of expert effort per chemistry is reduced to a single forward pass, yielding a universal polarizable force field that transfers across solvents, salts and additives. Coupled to automated system building and trajectory analysis, the platform reproduces experimental electrolyte properties with high fidelity: density error below 0.02 g/cm3, lithium-electrolyte ionic conductivity of approximately 1.5 mS/cm in quantitative agreement with measurement, and salt/additive solubility predicted to within 10%. We further demonstrate extended capabilities, including additive solubility by thermodynamic integration and electric-double-layer simulation under applied potential.
Nature Catal.
Desolvated cations promote CO2 electroreduction through partial covalent interactions
Yu Shan, Maximilian Jaugstetter, Julian Feijóo, and 14 more authors
Nature Catalysis, Sep 2026
Yu Shan, Maximilian Jaugstetter, Julian Feijóo and Yani Guan contributed equally.
Cations play a critical role in electrochemical CO2 reduction, yet the mechanisms underlying their influence remain incompletely understood owing to limited knowledge of their solvation structures and interfacial interactions under operating conditions. Here we combine surface-sensitive total electron yield X-ray absorption spectroscopy, surface-enhanced infrared absorption spectroscopy and density functional theory calculations to determine the solvation environment of caesium ions near a ligand-modified silver nanocatalyst under CO2-reducing conditions. We find that desolvated Cs+ ions, confined between a detached ligand layer and the silver surface, enable CO2 activation at potentials as high as 0.4 V versus reversible hydrogen electrode—reducing the overpotential for CO2-to-CO conversion by 250 mV compared with a bare Ag film. Our results further reveal partial covalent character in Cs–intermediate interactions, which directly accounts for the enhanced catalytic activity. These findings provide multimodal spectroscopic evidence that cations can modulate reaction kinetics beyond purely electrostatic effects.
Ag Amination
Electrochemical Amination of Acetone using Ag as Cathode: Mechanism and Role of Pb Impurities for Hydrogen Transfer
Yani Guan, Justus Kümper, Angelina Cuomo, and 6 more authors