OC20 IS2RE Adsorption Energy Prediction

What this research found

Computing how strongly a molecule binds to a catalyst surface normally requires hours of quantum-mechanical geometry relaxation per system. The Open Catalyst 2020 benchmark asks whether a model can skip that step and predict the relaxed energy from the unrelaxed starting geometry alone. Instead of a deep graph network, K-Dense built a deliberately transparent baseline: 60 hand-crafted descriptors fed to a gradient-boosting regressor, evaluated on all four official validation splits. Errors ranged from 0.762 to 0.918 electronvolts and reproduced the expected ordering, with unseen adsorbate chemistries proving hardest.

• Mean absolute errors across the four validation splits were 0.7725 electronvolts in-distribution, 0.9181 for unseen adsorbates, 0.7617 for unseen catalysts, and 0.8778 when both were unseen, averaging 0.8326 across roughly 25,000 systems per split. Under the stringent 0.02-electronvolt tolerance, only 2.44% of in-distribution predictions counted as essentially exact.
• Unseen catalysts were essentially free: the out-of-distribution catalyst error of 0.7617 electronvolts is marginally better than the in-distribution 0.7725, because a new alloy still sits in a smoothly varying space of electronegativity, radius, valence, and cell geometry that the training catalysts sample densely.
• Unseen adsorbates were the real difficulty, since adsorbate identity sets the zeroth-order binding strength while the catalyst only modulates it. The model's single most important feature was the count of carbon atoms in the adsorbate.
• Feature importance recovered the quantities that surface-science theory identifies as controlling chemisorption without being told to: binding-site electronegativity, covalent radius, and valence count alongside minimum adsorbate-surface distance and adsorbate height.
• The model trains in roughly 85 seconds on 16 threads with no graphics processor, against published graph-network errors of roughly 0.4 to 0.5 electronvolts, making it usable as a near-free pre-filter that reserves expensive relaxations for promising candidates. That gap comes from the fingerprint being a fixed, lossy summary where a graph network sees every atom and bond.

How it was done

The official benchmark archive was downloaded, verified against its published checksum, and read directly from its databases, with a sample of systems per split unpickled to confirm structures and targets were sound. For each system, 60 descriptors were computed from initial atomic positions, atomic numbers, unit cell, and per-atom surface tags only: 25 compositional, 14 geometric, 7 covering adsorbate-surface contact geometry with exact minimum-image treatment of periodic boundaries, and 14 chemical descriptors including binding-site properties. Relaxed coordinates were never read as inputs, guaranteeing the features are computable at inference time. A gradient-boosting regressor was trained on the provided 100,000-system subset under a mean-absolute-error objective matching the evaluation metric, split 90,000 for fitting and 10,000 for early stopping, converging at iteration 3,321 with a local validation error of 0.7616 electronvolts. The four official validation splits were untouched during fitting.

Data sources

• Open Catalyst 2020 initial-structure-to-relaxed-energy distribution — 100,000-system training subset plus four validation splits of roughly 25,000 systems each
• Chanussot et al., ACS Catalysis 11:6059-6072 (2021) — the benchmark itself, over 1.28 million relaxations across roughly 55 elements and 82 adsorbates
• Materials Project — source of the bulk metal and alloy slabs underlying the benchmark
• Atomic Simulation Environment and mendeleev — element-level covalent radii, masses, electronegativities, and valence counts