Letting the Surface Move: Lattice-Registered Induced Fit for Adsorbate Sampling

When a molecule lands on a crystal surface, the surface moves too. The top atomic layers buckle, shift, or pucker around the new bond, and that response can change which binding site is most stable. Current machine-learning samplers for molecules on catalyst surfaces hold the crystal fixed while they search for binding poses. Any surface response is added afterward by local energy minimization. Structural biology hit the same problem and addressed it with "induced fit" models, which reshape the protein pocket while placing the drug.

This note proposes carrying that idea to periodic crystals under the constraint that defines the semiconductor side of this pairing: lattice matching. The surface may relax only in ways that keep its atoms registered to the fixed bulk crystal underneath. That rule turns an open-ended, many-atom problem into a small set of periodic displacement patterns the sampler can move along. The same rule also lets us impose a deliberate mismatch, a uniform in-plane stretch or squeeze inherited from an underlying substrate, and ask how that strain and the induced fit interact.

The proposal has two parts, and the first is a measurement that could make the second unnecessary. We first measure how often "place on a rigid surface, then relax" lands in the wrong energy basin compared with searching over surface and molecule jointly. If that rate is negligible, the honest conclusion is that post-hoc relaxation suffices, and we would report that.

Core Claim

Our claim concerns adsorbates, meaning molecules bound to a surface. We claim that a meaningful fraction of their lowest-energy configurations sit in basins that a rigid-surface sampler cannot reach, even after local relaxation, because the surface's response decides which site wins. We further claim that this fraction grows when the surface carries epitaxial strain of about 1–2%. Epitaxial strain here means the stretch a thin film inherits when grown on a substrate with a slightly different lattice spacing.

The method is a sampler guided by a machine-learned interatomic potential, a neural network trained to reproduce quantum-mechanical energies and forces. It moves the molecule freely. It moves the surface only through a periodic displacement field whose average over the cell is zero and whose amplitude decays toward the fixed bulk layers. We expect this sampler to recover those missed minima at a cost comparable to rigid-surface sampling. We also expect the registry constraint, not merely "letting surface atoms move," to be what keeps the search stable and cheap.

What Exists and What It Leaves Open

The closest work is AdsorbSample. It samples low-energy adsorbate configurations on periodic catalyst slabs by conditional Boltzmann sampling, which means drawing configurations in proportion to how low their energy is. The sampling is guided by a pre-trained interatomic potential, with restraints against desorption and clashes. It lets the adsorbate bend internally but holds every slab atom fixed. It supplies our guidance machinery, and its fixed slab is the assumption we relax.

MatDock does flow matching, a generative method that learns a velocity field carrying random starting points to data, for several molecules inside periodic zeolite frameworks. It also treats the host as completely rigid. It motivates the gap but constrains ambitions: its roughly 10% success on unseen zeolites shows that learned vector fields for periodic hosts are still fragile.

On the biological side, three works supply the design vocabulary. FlexSBDD co-generates ligands while deforming protein backbones and sidechains, and reports that sidechain flexibility matters more than backbone motion. DynamicFlow runs a full-atom stochastic flow from predicted unbound pockets to bound ones. Harmony represents periodic torsion angles by Fourier series so that high-frequency noise is damped early.

We take two things from these works. First, a Fourier parameterization, applied now to a displacement field on the periodic cell rather than to angles. Second, a testable analogue of the FlexSBDD finding: do short-wavelength surface modes, which play the role of sidechains, matter more than long-wavelength strain-like modes, which play the role of the backbone?

PoseBusters is the main cautionary result. It showed that learned docking without physics produces clashing, distorted structures. That is why our sampler is driven by a physical potential rather than a learned score alone.

The semiconductor papers frame the constraint rather than supply a method. Wang et al. and Liu et al. sidestep lattice matching with amorphous buffers or conformal liquids. Qiu et al. discuss van der Waals stacking as another escape. In each, mismatch is treated as something to avoid. We treat a small, controlled mismatch as a knob whose effect on surface chemistry is jointly determined by the molecule and the relaxing surface.

We deliberately leave the wafer-scale topography work out of scope; its millimetre planarization lengths (Hymes et al.) have no honest link to a single unit cell.

Research Program

Surface parameterization. A slab is a few-layer periodic crystal model of a surface. Let $r$ be an atom's reference position in the slab, $z$ its height above the fixed bulk layers, and $G$ the reciprocal-lattice vectors of the supercell. The surface displacement field is

$$ u(r) = \sum_{G \neq 0} a_G \, e^{i G \cdot r} \, w(z), $$

where $a_G$ are learned complex amplitudes and $w(z)$ is a profile that is zero at the fixed bulk layers and one at the top layer. Excluding $G=0$ enforces lattice registry: the surface cannot slide or change its average spacing relative to the bulk. Epitaxial strain enters separately as a fixed biaxial strain $\epsilon$ applied to the whole cell. This treats the mismatch as a boundary condition, not a free variable.

Sampling objective. The sampler targets

$$ p(x, a) \propto \exp\!\left(-E_\theta(x, u_a; \epsilon)/k_B T\right), $$

where $x$ are the adsorbate coordinates, $u_a$ is the displacement field defined by amplitudes $a$, $E_\theta$ is the energy from a pre-trained MACE potential (acesuit/mace), and $k_B T$ is a temperature that we anneal. We reuse AdsorbSample's typed-branch guidance and restraints for the adsorbate. The noise schedule for surface modes follows Harmony's logic: high-$G$ modes are damped early, so the molecule commits to a region before the surface fine-tunes around it.

Diagnostic analysis. Before building anything we run the measurement. For each surface–adsorbate pair we compute a basin-switch rate. This is the fraction of pairs whose best joint-relaxation minimum is not reachable by full-slab local relaxation from any of the top rigid-surface samples. We compute it on unstrained slabs and again on slabs under ±1% and ±2% biaxial strain.

Baselines

The naive reference is heuristic site enumeration on the rigid slab followed by full MLIP relaxation of slab plus adsorbate. It shows how much any learned sampler adds.

The supervised reference is AdsorbDiff. On OC20-Dense in-distribution with 10 samples, AdsorbDiff reaches a 27.3% success rate (within 0.1 eV of the global minimum) and 8–13% diversity, as reported in AdsorbSample.

The strongest existing method is AdsorbSample itself. It reports 65.9% success and about 48–50% diversity on OC20-Dense in-distribution, and 60% success out-of-distribution, all with 10 samples. We will rerun it with the same MACE checkpoint rather than rely on the reported numbers. We will also run it with an added full-slab MLIP relaxation of every sample. This rigid-then-relax variant is the real competitor. The comparison is fair because all methods receive the same budget, counted in potential-energy force calls, and the same number of samples.

Beating it means two things. First, a statistically clear gain in 10-sample success on the strained splits. Second, at least parity on unstrained OC20-Dense at equal force-call budget.

Two ablations isolate the key ingredient. The first frees every top-layer atom independently, with no Fourier or registry constraint. It tests whether the constraint, not mere flexibility, carries the benefit. The second freezes either the high-$G$ or the low-$G$ modes, to test the sidechain-versus-backbone analogy.

Why not just…?

Why not just sample on a rigid slab and relax afterward? This is the most serious alternative. For many unstrained metal surfaces it may be enough. The biological evidence is suggestive but not decisive for crystals: FlexSBDD reports 6.43 average clashes for a rigid-pocket baseline against 1.39 with joint modeling, and shows cryptic pockets that rigid search never finds. Surfaces are stiffer than proteins, so the basin-switch diagnostic settles the question directly. If rigid-then-relax matches our success rate on strained slabs as well, we will say so. The contribution then reduces to the diagnostic and the strained benchmark.

Why not just unfreeze the surface atoms inside AdsorbSample's guidance? This is the cheapest version of induced fit, and it is our first ablation. The risk we expect is that per-atom freedom lets the sampler lower energy by distorting or partly reconstructing the slab. That would inflate the search space and the force-call budget. Whether registry-constrained modes are more sample-efficient is an empirical question. The ablation answers it at matched budget.

Why not just run basin hopping or molecular dynamics with the interatomic potential? Given enough compute, these search methods will find the joint minima. They therefore also help us build reference answers on the strained splits. The question is cost. DynamicFlow argues that explicit dynamics is impractical at screening scale, and we will check this by plotting success against force calls for basin hopping alongside our sampler.

Proposed Experiments and Decisive Checks

The primary benchmark is OC20-Dense, in-distribution and out-of-distribution, using the same 0.1 eV success criterion and diversity score as AdsorbSample. The new benchmark is a strained derivative of it. We apply ±1% and ±2% biaxial strain to OC20-Dense slabs. Reference minima come from exhaustive MLIP basin hopping, with DFT recomputation on a random subset of a few hundred pairs to check that MLIP errors do not decide the rankings. No reported result exists for any method on such a split, so every number there will be new and we will release the split.

The first decisive check is the basin-switch rate. If it stays near zero at every strain level, the joint sampler is unjustified and the paper becomes a negative result about rigid-slab sufficiency. The second is whether our sampler beats rigid-then-relax AdsorbSample on strained slabs at equal force-call budget without losing ground on unstrained OC20-Dense. The third is whether the registry-constrained ablation outperforms per-atom freedom. The fourth is which wavelength band carries the gain.

A secondary transfer check applies the same displacement field to zeolite hosts on the VOID Test 3 split, where MatDock reports 10.31% success with Voronoi-Gaussian sampling. It asks whether host flexibility reduces the framework clashes that dominate there. We treat it as exploratory, not as a headline result.