Does Low-Bit Quantization Erase the Weak Forces That Hold Molecular Crystals Together?

Neural network interatomic potentials are models that predict the energy of an arrangement of atoms, and they take the forces as the gradient of that energy. They have become accurate enough to replace expensive quantum chemistry when simulating organic molecular crystals. The packing of such a crystal is set by weak intermolecular forces: hydrogen bonds, the even weaker C–H···O contacts, and London dispersion. These contribute a few to a few tens of kJ/mol per molecule. The covalent bonds inside each molecule contribute far more. Quantization stores a network's weights in 2–4 bits instead of 16, and it adds rounding noise. In language models, this noise is known to wipe out small but important signals long before headline metrics move.

We propose to test a specific worry. A quantized potential could keep its average force error almost unchanged while losing the intermolecular part of the energy, which is the part that decides lattice energies, unit-cell volumes, and which polymorph is most stable. We will also build a quantization procedure aimed at that part. The procedure rests on two changes. First, sensitivity is measured with respect to the lattice energy, which is the crystal energy minus the energy of its isolated molecules. Second, force errors are split into whole-molecule motions and internal deformations. If the worry turns out to be correct, it also means that average force error is an unsafe acceptance test for any compressed potential.

Core Claim

Our hypothesis is that weight quantization of equivariant potentials degrades intermolecular quantities much faster than it degrades aggregate force error. An equivariant potential is one whose outputs rotate consistently when the input structure is rotated. The intermolecular quantities are lattice energy, the rigid-body forces and torques acting on each molecule, and polymorph energy ranking.

The proposed mechanism is the physical analogue of the outlier problem known from language models. Messages passed along long intermolecular edges, typically 2–6 Å and attenuated near the cutoff, are small in magnitude. Messages along short covalent edges are large, and they set the quantization scale. Inside layer-wise calibration, the small messages are treated as noise. We claim that calibrating and allocating precision against an intermolecular objective recovers lattice energetics at 3 and 2 bits, where force-calibrated quantization does not.

Why this is open, and what the closest work does

Two papers come closest, and neither tests this. FlashSchNet speeds up coarse-grained protein dynamics with fused kernels and channel-wise 16-bit quantization, staying within 0.04 GDT-TS of full precision. It never goes below 16 bits, never touches crystals or intermolecular energetics, and explicitly leaves equivariant architectures open. OMC25 builds 27 million PBE+D3-labelled molecular-crystal structures and trains EquiformerV2, eSEN and UMA. It shows that energy-conserving models and off-equilibrium crystal data are needed for reliable relaxations, but it evaluates only full precision. We take the OMC25 models and benchmarks as the testbed. We take the regime FlashSchNet avoided as the question.

Chemistry papers in the file motivate why small energies matter. The polysaccharide study The major role of London dispersion interaction in the assembly of cellulose, chitin, and chitosan decomposes cohesion and finds two things. Dispersion supplies 48–58% of cohesive energy, and a single hydrogen bond is worth at most 24–31 kJ/mol. The structural architecture of these crystals therefore rests on terms of tens of kJ/mol. The Solid‐State Structures of Dimethylzinc and Diethylzinc reports a cohesive energy of 0.61 eV/molecule for α-Me₂Zn, 94% of it dispersion. It also reports that the relative stability of its polymorphs is sensitive to small changes in dispersion parameters. That is exactly the kind of perturbation quantization noise could mimic. 1-Methyl-5-nitro-1H-imidazole shows that individually negligible C–H···O and C–H···N contacts repeat across the lattice and stabilize the packing. Its lesson is that errors on many weak contacts accumulate rather than cancel.

Quantization papers supply the methods and the cautionary results.

  • QuIP spreads outliers with random orthogonal rotations. It reaches 6.33 WikiText-2 perplexity at 2 bits on Llama 2 70B, where OPTQ collapses to 123.91. It optimizes each layer in isolation.
  • Cherry on Top finds that under 1% of parameters carry 50–300× higher sensitivity, and keeping them in 16-bit prevents 2-bit collapse.
  • CRVQ does the same at the channel level with extra codebooks.
  • GPTQT quantizes in two stages.

All four measure sensitivity against a single aggregate loss. Our departure is to change what sensitivity is measured against, not the quantizer. The finding that reasoning collapses faster than perplexity (Evaluating Quantized Large Language Models) is the language-model precedent for the claim that aggregate metrics hide selective loss.

Research Program

Restricting to weight-only quantization. We study weights at 8, 4, 3 and 2 bits with 16-bit activations. Weight-only quantization keeps the energy a smooth function of atomic positions, because the weights are constants. Forces obtained by differentiating the energy therefore stay exactly conservative. Activation rounding would make the energy piecewise constant in position and corrupt the forces, which would break the property OMC25 showed is necessary. We include one activation-quantized control only to show this.

The diagnostic. For each crystal we compare the quantized model with the full-precision model and with DFT on four quantities:

  • total force error;
  • force error projected onto each molecule's rigid-body subspace (net force and torque), versus the residual internal-deformation error;
  • lattice energy, defined as crystal energy per molecule minus the relaxed isolated-molecule energy;
  • the within-molecule energy ranking of the many randomly packed OMC25 structures sharing one molecule.

The selectivity hypothesis predicts that, as bits drop, the rigid-body and lattice-energy errors grow faster in relative terms than total force error.

The method. We replace the total-energy calibration objective with an intermolecular one:

$$\mathcal{L}_{\text{inter}} = \sum_{c}\Big(\Delta E^{q}_{c}-\Delta E^{\text{fp}}_{c}\Big)^2 + \lambda \sum_{c}\big\lVert P_{c}\,(F^{q}_{c}-F^{\text{fp}}_{c})\big\rVert^2 .$$

The terms are:

  • $c$ indexes calibration crystals;
  • $\Delta E_c$ is the lattice energy of crystal $c$ under the quantized ($q$) or full-precision ($\text{fp}$) model;
  • $F_c$ are atomic forces;
  • $P_c$ projects forces onto per-molecule rigid-body motions;
  • $\lambda$ balances the two terms.

We use $\mathcal{L}_{\text{inter}}$ in three places. It replaces the total loss when computing CherryQ-style squared-gradient sensitivity, which picks the roughly 0.4–1% of weights kept in 16-bit. It reweights calibration samples in the GPTQ/QuIP layer-wise proxy, so that layers are fitted on inputs that matter for intermolecular energy. And it drives a short end-to-end tuning of only the quantization scales, which addresses the inter-layer error accumulation that layer-wise methods ignore.

Preserving equivariance. Any QuIP-style rotation is restricted to mixing channels within the same angular-momentum degree. Mixing across degrees would make rounding errors break rotational equivariance.

Analysis. We will locate where the intermolecular-sensitive weights live, for example in radial filters for long distances or in particular attention heads. This answers whether potentials, like language models, concentrate critical function in a small fraction of parameters.

Baselines

The naive reference point is the full-precision model. We will use the released UMA-M-1.1 and EquiformerV2 OMC25 checkpoints and run them in their original precision and in BF16. The reported full-precision numbers are:

  • UMA-M-1.1 on X23b: lattice energy MAE of 1.94 kcal/mol and unit-cell volume error of 5.78% (OMC25);
  • EquiformerV2 on the OMC25 test set: force MAE of 3.79 meV/Å.

These define the ceiling. No reported result was found on Lacuna for any quantized interatomic potential on OMC25 or X23b. The low-bit comparisons are therefore all new measurements, and we will reproduce the full-precision numbers first.

The strongest existing method is QuIP-style rotation plus adaptive rounding, adapted with degree-preserving rotations. It is the method that most reliably avoids the low-bit cliff in the file. We pair it with CherryQ-style mixed precision whose sensitivity is computed from the standard total-energy-and-force loss. Running both rotation and mixed precision is the strongest combination available, and it is exactly what a practitioner would try first.

Plain GPTQ is included as the common default. Channel-wise uniform rounding, the FlashSchNet recipe pushed to 8 and 4 bits, is included as the direct extension of the closest existing work.

All methods share the following controls:

  • the same calibration set of OMC25 crystals plus their isolated molecules;
  • the same bit budget, counting any 16-bit weights;
  • the same budget for tuning the quantization scales;
  • the same hyperparameter search over $\lambda$, or over its analogue for the baselines.

Ablation of the key ingredient. This is the central control: we run identical pipelines in which the only change is whether sensitivity and calibration use $\mathcal{L}_{\text{inter}}$ or the standard loss.

Definition of beating the baseline. At 3 bits, X23b lattice energy MAE stays within 0.5 kcal/mol of the full-precision 1.94 kcal/mol while the strongest baseline does not. At 2 bits, polymorph ranking (Spearman correlation and top-1 agreement with the DFT minimum) is clearly better. In both cases, total force MAE should be no worse than the baseline's.

Why not just…?

Why not just stay at 16 bits, as FlashSchNet does? For many uses this is the right answer. FlashSchNet also showed that edge-tensor memory traffic, not weight storage, dominates cost in its setting. Low-bit weights may therefore buy little speed for small models, and we do not promise speedups. The payoff is in two places. One is memory for large universal models and many-replica screening. The other matters regardless of deployment: the diagnostic tells us whether aggregate force error can certify any compressed potential, whether it was quantized, distilled or pruned. If quantized models pass the diagnostic with standard methods, that is itself a useful negative result.

Why not just apply QuIP rotations, which already tame outliers? This is a real threat. If small intermolecular signals are lost only because large covalent messages inflate the scale, rotations alone may fix it. The rotation baseline settles this directly. If rotated GPTQ matches our method on lattice energy and ranking, the objective is unnecessary and we will report that. Our expectation is otherwise, for a specific reason. Rotations make each layer's error uniform across directions, but they do not tell the quantizer which outputs matter. An error budget spread evenly still lands on the small terms.

Why not just quantization-aware training on the standard force and energy loss? Retraining with rounding in the loop lets the whole network adapt. CRVQ reports more than 7,000 GPU hours for this on LLaMA2-13B versus 50 for its post-training method. Cherry on Top reports 9.25 C4 perplexity for uniform quantization-aware training at 3 bits versus 7.39 for its mixed-precision method. Interatomic potentials are much smaller, though, so cost is a weaker argument here. We will run quantization-aware training at matched GPU hours. If it recovers lattice energies without the intermolecular objective, the objective's value reduces to cheapness.

Why not just add more crystals to the calibration set? The issue might simply be calibration data drawn from the wrong distribution. Our ablation separates data from objective by including the same crystals and isolated molecules with the standard loss.

Proposed Experiments and Decisive Checks

All experiments use released OMC25 checkpoints, the OMC25 test split, and X23b.

First check: selectivity. We plot relative degradation of rigid-body force error, lattice energy and polymorph ranking against relative degradation of total force error, across bit widths and baselines. The hypothesis is falsified if these move together. In that case we would report the result as reassurance that standard quantization is safe for crystals.

Second check: relaxations. We relax X23b and a held-out OMC25 subset with each quantized model. We report:

  • unit-cell volume error against the full-precision 5.78%;
  • the fraction of relaxations ending in the same structure as the full-precision model;
  • the rate of unphysical collapse, the failure mode OMC25 associated with poorly trained potentials.

Third check: conservation. We run short constant-energy molecular dynamics on several crystals. We compare energy drift for weight-only models against a model with 8-bit activations. The expectation is no added drift for weight-only models and measurable drift once activations are rounded.

Fourth check: transfer of the sensitive weights. We test whether the weights flagged by the intermolecular objective in UMA also matter in EquiformerV2's corresponding modules. As an optional out-of-distribution stress test, we will recompute PBE+D3 inter-chain energies for cellulose Iβ, for which the polysaccharide study reports 62 kJ/mol dispersion with D2. We then ask whether quantized models preserve the inter-chain ordering. This test is informative only if the full-precision model already gets it right.