MAD-SURF: a machine learning interatomic potential for molecular adsorption on coinage metal surfaces
Manuel González Lastre, Joakim S. Jestilä, Rubén Pérez, Adam S. Foster · 2026
Read MAD-SURF: a machine learning interatomic potential for molecular adsorption on coinage metal surfaces by Manuel González Lastre, Joakim S. Jestilä, Rubén Pérez, Adam S. Foster (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-18852.txt.
Introduction
Manuel González Lastre and three colleagues build a fast stand-in for the strict maths of how bits of chem work stick to gold-like coin metals. How such bits land, group, and pull on such a face is key to new chips and to how we read fine shots of that face. The strict maths is too slow for the big counts of atoms in a real face. Teams need a tool that is near as true but far more swift.
Their tool learns the push and pull from a set of strict runs. Once taught, it can run big faces at a small share of the cost. A swift, true tool could let teams screen many bits for a face in a day. It could also help read the fine shots that new probes take of such faces.
Catalog record
| CREATOR | Manuel González Lastre, Joakim S. Jestilä, Rubén Pérez, Adam S. Foster |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-18852
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MAD-SURF: a machine learning interatomic potential for molecular adsorption on coinage metal surfaces Manuel González Lastre, Joakim S. Jestilä, Rubén Pérez, Adam S. Foster Abstract Predicting how organic molecules adsorb, assemble, and interact on metal surfaces is central to surface chemistry and molecular electronics, particularly in the context of interpreting high-resolution scanning probe microscopy. Yet, the application of first-principles simulations to interfaces is hampered by the computational cost for evaluating the electronic structure for the large number of atoms typically involved. We hereby present MAD-SURF, a machine learning interatomic potential specifically tailored for molecular adsorption on coinage metal surfaces. Trained on a broad dataset spanning diverse molecules, adsorption motifs, surfaces, molecular dynamics trajectories and non-covalent aggregates, MAD-SURF achieves accuracy comparable to the underlying DFT reference while enabling simulations orders of magnitude faster than density functional theory. The model reliably reproduces energies, forces and adsorption geometries across the three coinage metal substrates. We demonstrate its capabilities on experimentally characterized systems, including organic monolayers, polycyclic aggregates, flexible biomolecules and the long-range herringbone reconstruction of gold. By merging accuracy, speed, and generalizability, MAD-SURF offers a practical framework for accelerating atomistic simulations and advancing data-driven workflows in surface science. † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. Keywords: machine learning interatomic potentials, surface chemistry, molecular dynamics, materials modeling, density functional theory, scanning probe microscopy Introduction Metal-organic interfaces are central to modern materials science Liu2024Mar ; Scott2003Mar , providing the structural and electronic foundations for applications in catalysis norskov2009towards , molecular electronics Heimel2007Apr ; Vuillaume2003Nov ; Dasari2011Jun , and nanofabrication Lugier2021Sep ; Liu2017 . Density Functional Theory (DFT) is currently the main workhorse for modeling such interfaces, but its computational cost generally scales near cubically with system size, i.e. \mathcal{O}(N^{3}) , due to the diagonalization of the Kohn-Sham Hamiltonian. In practice, this limits calculations from a few hundred up to around a thousand atoms even on modern supercomputers keller2024small . For more complex systems, such as large hydrocarbons in petroleum Zhang2021Sep ; chen_gross_acs_fuels , molecular monolayers Gardener2009Dec ; Galeotti2020Aug ; DeMarchi2019 , biomolecules Grabarics2024Nov ; Cai2025Mar ; jestila_lignocellulosic ; milica_stradiol , or DNA Pawlak2019Feb ; Cai2022Nov , the cubic scaling of DFT quickly renders it prohibitively costly, necessitating the development of faster and more scalable approaches while retaining its accuracy maurer_advances_2019 . Recently, machine learning interatomic potentials (MLIP) have emerged as a powerful alternative that combines near-DFT accuracy with orders-of-magnitude lower computational cost. They have successfully been applied across diverse domains of chemistry and materials science, from heterogeneous catalysis oc2020 ; oc2022 ; adsorbml , to biomolecular simulation Noe2020Apr ; microsoft_ai2bmd , and even the discovery of new materials mattersim ; deep_mind_materials_discovery_nature . Despite these advances, most existing efforts have concentrated on crystalline solids m3gnet_nat_comp_sci ; mattersim ; chgnet_nat_mach_int ; deep_mind_materials_discovery_nature , reactive adsorbates oc2020 ; oc2022 ; stark_benchmarking_2024 or biomolecular systems in solution Noe2020Apr ; microsoft_ai2bmd ; spice_nat_sci_data ; ani_ml_jctc . Notably, foundation models covering a wide array of materials including elements from most of the periodic table have recently emerged, providing pretrained potentials that can be finetuned to provide accurate properties for any system of interest with a much smaller dataset than when training from scratch chgnet_nat_mach_int ; batatia_foundation_2025 ; pet_mad_nat_comm . However, these models are mainly trained on bulk materials, and analogous foundation models explicitly targeting dispersion-dominated, surface-screened adsorption and collective assembly of large organic adsorbates are currently lacking. Furthermore, polycyclic adsorbates stabilized primarily through non-covalent interactions at metal surfaces Liu2014Nov remain underrepresented in current MLIP benchmarks and training sets westermayr_long-range_2022 , despite their importance in modern materials science. In this context, On-Surface Synthesis grill_nano-architectures_2007 ; clair_controlling_2019 combined with Scanning Probe Microscopy (SPM) has emerged as a unique tool to investigate these systems, enabling direct visualization of adsorption geometries, charge transfer, and even reactive intermediates at the atomic scale pavlicek2017 ; patera_probing_2025 . This capability has enabled detailed studies of catalytic reaction pathways AlbrechtJACS2015 ; PavlicekNatChem2015 , single-molecule charge transport in molecular electronics Bumm1996Mar ; Fatayer2018May , and biomolecular recognition at interfaces Hinterdorfer2006May ; Kienberger2006Jan , providing insights not accessible to ensemble measurements. Nonetheless, despite the availability of efficient and accurate simulation tools ellner2019molecular ; niko_oinonen_ppm_review , interpretation for more complex metal-organic interfaces where van der Waals forces and conformational flexibility govern adsorption geometries and image contrast is still constrained by the limitations of DFT. Accelerating the modeling of such systems while retaining near-DFT accuracy is essential for large-scale screening of potential adsorbate configurations todorovic_bayesian_2019 ; egger_charge_2020 ; jestila_lignocellulosic , in automating experimental workflows for molecular reactions nian_jacs_spm and discovery Kalinin2021Aug ; Li2025Jul , as well as in terms of democratizing the computational sciences for researchers without access to extensive computational resources or high-performance computing facilities. Furthermore, recent work has shown that MLIPs can resolve long-standing problems in surface science out of scope for even DFT, in particular large-scale surface reconstruction phenomena. To this end, Li and Ding were able to explain the atomic-scale origin, thickness dependence and strain sensitivity of the Au(111) herringbone reconstruction with a DeePMD potential trained specifically for the Au surface, a system that is inaccessible to conventional ab initio simulations owing to its large periodicity li_origin_2022 . Accounting for such ubiquitous surface reconstructions in simulations provides inevitably more realistic models for catalysis chen_machine-learning_2023 , self-assembly sampath_closing_2020 and thin-film deposition pieck_computational_2025 , and in combination with the extended time and length-scales at near-DFT accuracy, MLIPs represent a transformative tool for realistic large-scale MD simulations. Still, achieving near first-principles accuracy while simultaneously accessing the time and length scales required to describe adsorption dynamics, collective assembly, and thermal fluctuations at metal–organic interfaces remains a central challenge, particularly for large, flexible organic adsorbates stabilized by dispersion-dominated interactions on metal surfaces. In this work, we introduce MAD-SURF (Molecular ADsorption on SURFaces), a machine learning interatomic potential tailored for molecular adsorption on coinage metal surfaces. MAD-SURF is trained on a diverse dataset specifically constructed for interfacial and on-surface chemistry, with all reference calculations performed at the Density Functional Theory level (PBE+vdW {}^{\text{surf}} ). We focus on the widely used coinage metal surfaces (Cu, Ag, and Au), owing to their well-characterized electronic properties and distinct adsorption strengths. The dataset combines adsorption configurations obtained through active learning-based structure search, ab initio molecular dynamics trajectories of molecules on these surfaces, normal mode sampling to capture intramolecular flexibility, and automated sampling of molecular clusters relevant to intermolecular interactions in aggregation and monolayer formation. This multi-source strategy systematically covers local bonding, long-range dispersion, and conformational degrees of freedom, providing a robust foundation for training interatomic potentials for on surface chemistry. Using this dataset, we benchmark different training strategies for the state-of-the-art MACE batatia2022mace architecture on their accuracy to predict energies, forces, adsorption heights, overall geometries, and transferability to larger adsorbates. From this comparison, the best-performing potential is selected, enabling simulations of extended organic adsorbates on surfaces with accuracy comparable to our DFT reference at a fraction of the computational cost. The resulting MAD-SURF potential provides a practical framework for simulating the adsorption, assembly, and dynamics of organic molecules on coinage metal surfaces with near first-principles accuracy. Beyond adsorption of isolated molecules, we demonstrate its generalization across chemically diverse systems, including organic monolayers, polycyclic molecular aggregates, and