AI model built on copper quickly learned nickel CO2 reactions, with errors up to 0.03 eV

Study title: Carbon dioxide hydrogenation on copper and nickel catalysts <i>via</i> a conformal sampling approach

A peer-reviewed study introduced the Conformal Sampling of Catalytic Processes workflow to model heterogeneous CO2 hydrogenation on copper and nickel. Starting from exhaustive DFT on Cu(100), the team trained a Machine Learning Interatomic Potential and transferred it to Ni(100) to test cross-metal portability.

Read the study abstract

We present a computational workflow, the Conformal Sampling of Catalytic Processes (CSCP) approach, and its application to the case of heterogeneous hydrogenation/reduction of carbon dioxide (CO 2 ) on copper and nickel catalysts. CO 2 activation is of critical importance for a sustainable global future and one of the major reactions for which sustainable routes must be found urgently. We use the fcc(100) facet of Cu as a worked-out case, and exhaustively derive at the DFT level its reaction mechanisms. We then…

By the Truvace deskSeptember 17, 20261 min read1 sourceEvidence: Moderate · 52/100 — Benchmark or lab evaluation · 1 independent sourcePeer-reviewedTRV-2026-1119· cite & verify

Carbon dioxide hydrogenation on copper and nickel catalysts <i>via</i> a conformal sampling approach — "Has the world changed and we haven't noticed? #renewable #energy @RadioNational" by planeta is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

The result matters for climate mitigation because accurate, transferable MLIPs could accelerate rational design of catalysts for sustainable CO2 conversion. Uncertainty remains about how the observed pathway divergences in exception cases affect predictions on other facets, alloys, or operating conditions beyond the two monometallic (100) surfaces tested.

Main points

Workflow named Conformal Sampling of Catalytic Processes (CSCP) was applied to heterogeneous hydrogenation of CO2 on Cu and Ni.
Initial exhaustive DFT mechanisms were derived on fcc(100) facet of Cu, then knowledge was carried over to Ni(100) to test transferability.
MLIPs maintained uniform accuracy across complex reaction diagram and explored alternative atomistic pathways in exception cases.

The gain

In computer benchmarks, AI models of CO2 hydrogenation had maximum errors of 0.05 eV on copper and 0.03 eV on nickel, and the copper work was carried over to nickel rapidly.

The rundown

Researchers built an exhaustive DFT reaction network for CO2 reduction on Cu(100), then used CSCP to train an MLIP and transferred that knowledge to Ni(100) to rapidly derive a second MLIP.

Evaluation showed consistent prediction of energies and barriers across all mechanistic steps, with larger discrepancies explained by the MLIP finding alternative pathways rather than loss of accuracy, supporting use for catalyst rational design.