MAD-SURF: a machine learning interatomic potential for molecular adsorption on coinage metal surfaces

The paper introduces MAD-SURF, a machine learning interatomic potential trained on diverse datasets that achieves density functional theory-level accuracy for simulating organic molecule adsorption on coinage metal surfaces while enabling computational speeds orders of magnitude faster.

The Gist

Imagine you are trying to understand how a specific type of Lego brick (an organic molecule) sticks to a shiny metal table (a coin like gold, silver, or copper). This is a huge deal for scientists because it helps us build better medicines, faster computers, and new materials.

The problem is that figuring out exactly how these Lego bricks stick is incredibly hard.

The Old Way: The Slow, Expensive Calculator

For years, scientists used a method called Density Functional Theory (DFT). Think of DFT as a super-precise, high-resolution 3D scanner. It can tell you exactly how every single atom in the Lego brick and the metal table interacts. It's accurate, but it's also painfully slow.

If you want to scan one brick, it takes a few minutes. But if you want to scan a whole city of bricks moving around, interacting, and rearranging themselves over time? That calculation would take a supercomputer years to finish. It's like trying to paint a masterpiece by calculating the exact color of every single pixel one by one; the result is perfect, but you'll never finish the painting.

The New Solution: MAD-SURF (The "Smart Guess" Engine)

The authors of this paper created a new tool called MAD-SURF.

Think of MAD-SURF as a brilliant apprentice who has studied the master scanner (DFT) for years.

1. The Training: The researchers fed the apprentice millions of examples of Lego bricks sticking to metal tables. They showed it every possible angle, every type of brick, and every way they could wiggle or stick together.
2. The Learning: Instead of doing the heavy math for every single new situation, the apprentice learned the patterns. It learned, "Oh, when a brick with a nitrogen atom touches gold, it sits at this height. When two bricks touch, they lean this way."
3. The Result: Now, when you ask MAD-SURF to simulate a million bricks moving around, it doesn't do the heavy math. It makes a "smart guess" based on what it learned.

The Magic: MAD-SURF is thousands of times faster than the old scanner, but it's still almost as accurate. It's like the apprentice can paint the whole city in an hour, and the picture looks just as good as the one the master took years to paint.

What Did They Test?

To prove their apprentice was actually good, they tested it on some very tricky scenarios:

• The "Petroleum" Puzzle: They looked at complex, messy piles of oil molecules on copper. The old method was too slow to see how they settled. MAD-SURF quickly figured out the exact shape, matching real-life microscope photos perfectly.
• The "Carpet" Test: They simulated a whole floor covered in organic molecules (like a carpet). They checked if the molecules stood up straight or lay flat. MAD-SURF got the spacing and height almost exactly right, just like the real experiments.
• The "Living" Test: They tried simulating a flexible biological molecule (a sugar ring called cyclodextrin). These molecules wiggle and twist. MAD-SURF handled the wiggling and showed how the molecule sits on a gold surface, matching what scientists see under microscopes.
• The "Gold Skin" Test: Gold surfaces have a weird, wavy pattern (like a herringbone fish skeleton) that is hard to simulate because it requires a huge area. Usually, computers crash trying to simulate this. MAD-SURF handled the massive size and reproduced the wavy pattern perfectly, even though it wasn't specifically taught to do just that one thing.

Why Does This Matter?

Before MAD-SURF, scientists were stuck. They could either:

1. Be accurate but only look at tiny, static snapshots (like a single photo).
2. Be fast but look at huge systems with low accuracy (like a blurry, low-res video).

MAD-SURF lets them have both. It allows scientists to watch movies of molecules dancing, sticking, and building themselves on metal surfaces in real-time, with near-perfect accuracy.

In short: MAD-SURF is a super-fast, super-smart simulator that lets us finally see the "movie" of how molecules build our future technology, without waiting years for the computer to finish the calculation.

Technical Summary

1. Problem Statement

The accurate simulation of organic molecules adsorbing, assembling, and interacting on metal surfaces is critical for surface chemistry, molecular electronics, and catalysis. However, current methodologies face a significant bottleneck:

• Density Functional Theory (DFT): While accurate, DFT scales cubically ($O(N^3)$) with system size due to Kohn-Sham Hamiltonian diagonalization. This limits simulations to a few hundred atoms, making it prohibitively expensive for large systems like molecular monolayers, polycyclic aggregates, biomolecules, or long-range surface reconstructions (e.g., Au(111) herringbone).
• Classical Force Fields: While computationally cheap, they lack the quantum mechanical accuracy required to describe complex bonding, charge transfer, and dispersion-dominated interactions accurately.
• Existing Machine Learning Interatomic Potentials (MLIPs): Most existing MLIPs are trained on bulk crystalline solids or reactive adsorbates. There is a lack of "foundation models" specifically tailored for dispersion-dominated, surface-screened adsorption and the collective assembly of large organic molecules on coinage metals (Cu, Ag, Au).

2. Methodology

A. Dataset Construction (MAD-SURF)

The authors constructed a comprehensive, chemically diverse dataset specifically for on-surface chemistry, calculated using DFT (PBE+vdWsurf) as the reference. The dataset spans 38 molecules containing key functional groups (aromatics, amines, carbonyls, halogens, etc.) adsorbed on Cu(111), Ag(111), and Au(111).
The data generation pipeline utilizes four complementary strategies to ensure coverage of local bonding, long-range dispersion, and conformational flexibility:

1. Bayesian Optimization Structure Search (BOSS): To find adsorption minima and explore the potential energy surface (PES).
2. Ab Initio Molecular Dynamics (AIMD): To capture thermal fluctuations and dynamic behavior of molecules on surfaces and bulk slabs.
3. Normal Mode Sampling (NMS): To perturb intramolecular degrees of freedom and capture flexibility.
4. Automated Interaction Site Screening (AISS): To sample non-covalent intermolecular interactions (dimers, clusters) relevant to aggregation.

The final dataset contains tens of thousands of configurations (energies and forces) covering gas-phase molecules, adsorbed states, and molecular aggregates.

B. Model Architecture and Training Strategy

The study utilizes the MACE (Machine Learning Atomic Cluster Expansion) architecture, a state-of-the-art equivariant neural network. The authors systematically benchmarked several training strategies:

• Training from Scratch: Using the full dataset or subsets with varying force/energy loss weight ratios ($\lambda$).
• Filtering: Naïve filtering (every $N$-th entry) and descriptor-based filtering (using MACE descriptors and FAISS-gpu to remove redundant configurations).
• Transfer Learning (Finetuning): Starting from the pretrained MACE-MPA-0 foundation model (trained on bulk materials) and finetuning it on the MAD-SURF dataset.
• Few-Shot Learning: Rapid adaptation for specific target molecules using a small set of static reference calculations.

Key Finding: Models trained from scratch failed to accurately predict adsorption heights despite low energy/force errors. Finetuning the MACE-MPA-0 foundation model yielded the best results, significantly improving adsorption height correlation ($R^2$) and structural fidelity.

3. Key Contributions

1. Development of MAD-SURF: A general-purpose MLIP specifically designed for molecular adsorption on coinage metals (Cu, Ag, Au).
2. Novel Training Protocol: Demonstration that finetuning a bulk-trained foundation model is superior to training from scratch for surface adsorption problems, particularly for capturing correct adsorption geometries.
3. Comprehensive Benchmarking: A rigorous evaluation showing that energy/force errors alone are insufficient metrics; adsorption height parity is a critical indicator of model quality for surface systems.
4. Open Resources: Release of the dataset, pretrained models, and code to accelerate research in surface science.

4. Results and Validation

The authors validated MAD-SURF across four distinct categories of systems, demonstrating transferability and accuracy comparable to DFT at a fraction of the computational cost:

• Molecular Aggregates: Successfully reconstructed atomic arrangements of thermally induced polycyclic aggregates (from 2,7-dimethylpyrene) on Cu(111). The simulated Non-Contact AFM (NC-AFM) images matched experimental contrast, validating the model's ability to reproduce molecule-molecule interactions.
• Organic Monolayers: Reproduced the structures of pentacene and PTCDA monolayers on Cu(111), Ag(111), and Ag(110).
  • Adsorption heights deviated by less than 0.2 Å from experimental values.
  • Captured the delicate balance between intermolecular hydrogen bonding and substrate interactions.
  • Noted that discrepancies in PTCDA/Ag(111) heights were inherited from the DFT reference limitations, confirming the model's fidelity to its training data.
• Biomolecules: Modeled the adsorption of $\beta$-cyclodextrin on Au(111). The model correctly resolved the hydrogen-bonding networks and distinguished between primary and secondary face orientations, matching experimental AFM features.
• Long-Range Surface Reconstruction: Successfully simulated the Au(111) herringbone reconstruction (a ~30 nm periodic pattern) without any Au-specific active learning.
  • The model captured the correct periodicity (302.9 Å vs. experimental ~324.3 Å) and surface corrugation (0.2 Å).
  • This proves the MLIP encodes the long-range elastic response of the metal substrate, a feature usually requiring specialized potentials.
• Large-Scale Dynamics: Performed molecular dynamics simulations on a 13.27 × 13.71 nm² system containing 118 pentacene molecules (17,144 atoms).
  • Observed the transition from a random-tiling phase to a bilayer structure at elevated temperatures (420 K), phenomena inaccessible to DFT due to system size and timescale limitations.

5. Significance

• Bridging Scales: MAD-SURF merges near-DFT accuracy with the scalability of classical MD, enabling simulations of systems with thousands of atoms over nanosecond timescales.
• Experimental Interpretation: It provides a practical framework for interpreting high-resolution Scanning Probe Microscopy (SPM) data by generating atomistic models that match experimental contrast.
• Accelerated Discovery: By lowering the computational barrier, it facilitates the screening of adsorbate configurations, the study of self-assembly mechanisms, and the exploration of complex metal-organic interfaces.
• Future Outlook: The framework sets a precedent for developing specialized MLIPs for on-surface synthesis, heterogeneous catalysis, and molecular nanofabrication, with potential extensions to reactive events and other substrates.