Machine-learned force fields predict the force on every atom of a molecule. Molecular dynamics simulations then call them millions of times in a row. That repeated inference, more than training, is where most of the compute goes. Spiking neural networks are networks whose neurons talk in binary pulses, so multiplications become cheap additions. They promise large energy savings, but they have mostly been shown on classification, and they are known to lose precision when asked for continuous outputs. This note proposes a way to decide where a low-precision force model must stay precise. The answer comes from work on anisotropic noise: the observation that atoms move easily along some directions (twisting a rotatable bond) and barely along others (stretching a covalent bond).
The core idea fits in two sentences. A force error along a stiff bond barely moves the atom, but the same error along a soft direction sends it drifting. The per-atom noise covariance that anisotropic denoising methods already learn is, to first order, a map of this softness. We therefore use that learned covariance to weight the training loss and to split a fixed spike (or bit) budget across directions, putting precision on soft modes and letting stiff modes be coarse. The test is whether, at equal estimated energy, this beats uniform precision on force accuracy and, more importantly, on how long a simulation runs before it fails.
Core Claim
Denoising pre-training perturbs atom positions and trains a network to undo the perturbation. The learned "score" then approximates physical forces. Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields (AniDS) generates a full, rotation-equivariant 3×3 covariance $\Sigma_i$ for each atom $i$. It suppresses noise along covalent bonds and links the denoising loss to the Boltzmann score. Under a local harmonic picture, $\Sigma_i \approx k_B T\, H_{ii}^{-1}$, where $H_{ii}$ is the local block of the energy Hessian and $k_B T$ is the thermal energy. A force error $\delta F_i$ then displaces the atom by roughly
$$\delta x_i \approx H_{ii}^{-1}\,\delta F_i \approx \Sigma_i\,\delta F_i / (k_B T),$$
so the error that matters for a simulation is the compliance-weighted error $\lVert \Sigma_i \delta F_i \rVert$, not the raw force error.
A rate-coded output with $T_d$ time steps along direction $d$ has resolution roughly proportional to the force range along $d$ divided by $T_d$. Take a harmonic mode with stiffness $k_d$. Its typical thermal force scales as $\sqrt{k_d k_B T}$, and its compliance is $1/k_d$. Minimizing the summed squared compliance-weighted error subject to a fixed total budget, via a Lagrange multiplier, gives
$$T_d \propto \left(k_B T / k_d\right)^{1/3}.$$
Soft modes therefore deserve more output precision than stiff ones, with a specific cube-root law. Nothing in the spiking literature or the anisotropic-denoising literature makes this prediction or tests it.
The claim is deliberately stated for low-precision force models in general, not only spiking ones. A reviewer could reasonably argue that integer quantization gets similar savings more easily. We therefore test the same allocation principle under both spiking and integer quantization, so the contribution survives even if spikes lose.
Research Program
The teacher is an AniDS model with an Equiformer backbone, pre-trained on PCQM4Mv2 using the official code (ZeroKnighting/AniDS) and fine-tuned on MD17. Its noise generator supplies $\Sigma_i$. Because a simulation keeps the bonding topology fixed, $\Sigma_i$ is computed once per molecule at equilibrium in full precision and reused for every step. Its cost is amortized away rather than paid per force call.
The student is a spiking network. It handles rotations by never spiking vectors: spikes carry only rotation-invariant quantities, namely distances, angles, and atom types. The force is assembled as $F_i = \sum_j c_{ij}\,\hat u_{ij}$, a sum of invariant scalar coefficients $c_{ij}$ times unit vectors $\hat u_{ij}$ pointing along neighbor edges. This construction is exactly equivariant. It isolates the continuous multiply to one small output stage, which we count in the energy budget. It also directly addresses a gap in 3DSMT, which encodes raw coordinates into spikes for point-cloud classification with no rotational equivariance.
Inside the student we reuse two design choices from Spiking Diffusion Models: residuals summed before the spiking nonlinearity, and learnable per-time-step input scaling. That paper showed these keep continuous outputs (image pixels) usable from spiking denoisers. We also distill from the continuous teacher, following the ANN-to-SNN distillation in SpikingBERT.
The proposal's own ingredient has two parts, both computed from $\hat u_{ij}^\top \Sigma_i \hat u_{ij}$, the compliance along each edge. The first is a training loss weighted by that compliance. The second is a per-edge output budget that follows the cube-root law. In the spiking version, the budget sets the number of output time steps or the width of the neuron population encoding each $c_{ij}$. In the quantized version, the same rule sets per-edge bit width.
A secondary analysis checks the premise directly. We compare the eigen-directions and eigenvalues of $\Sigma_i$ against the teacher's own energy Hessian, obtained by automatic differentiation, on MD17 conformations. Sliced Denoising reported a 42% gain in correlation between learned force fields and quantum-chemistry ground truth over isotropic denoising. That result suggests learned noise tracks physics, but it does not show that the covariance itself is a good compliance map. If the agreement is poor, the allocation should fall back on a fixed physical prior, which we also test.
Baselines
The naive reference point is a uniform-precision student: the same spiking architecture, the same total time steps, an unweighted force loss, and identical distillation. Every other comparison is measured against it at matched estimated energy.
The strongest existing accuracy reference is the continuous AniDS teacher itself. The file reports an 8.9% average relative improvement in MD17 force accuracy over leading isotropic denoising models for AniDS. It reports a 20.9% gain on Naphthalene for DenoiseVAE over Frad and SliDe. No absolute MD17 force mean absolute error was found on Lacuna for any of these methods, so we report the teacher's MAE as measured under our own fine-tuning. The student "beats" nothing here. The question is how close it comes per unit energy.
The strongest efficiency competitor is the teacher quantized to INT8 and INT4 with uniform bit width, tuned with the same budget of hyperparameter trials as the student. The energy reference points in the file are 0.9 pJ per accumulate versus 4.6 pJ per 32-bit multiply-accumulate in 45nm CMOS (SpikingBERT), and a roughly 14× ratio reported by Deep Spiking Neural Networks for Large Vocabulary Automatic Speech Recognition. No integer multiply-accumulate energy appears in the file. We will therefore state the per-operation table we adopt and report how the conclusions change across a range of plausible INT8 costs.
All models share data splits, the pre-trained teacher, and the training-epoch budget. Energy is estimated as operation counts times per-operation energy, the same method behind the 4.3 mJ versus 207.0 mJ comparison in 3DSMT. We state openly that these are estimates, not measurements on neuromorphic hardware.
Three ablations isolate the key ingredient:
- Compliance-weighted loss with uniform budget.
- Budget allocation without the weighted loss.
- Allocation from a fixed bond, angle, and torsion stiffness prior in the style of Sliced Denoising instead of the learned $\Sigma_i$.
Ablation 3 tests whether learning the covariance matters at all. The proposal counts as successful if compliance-weighted allocation meets either of two conditions against uniform allocation at the same estimated energy:
- It runs stable simulations significantly longer, or
- It reaches the same stability with clearly less energy.
The success also has to hold for both spiking and quantized students.
Why not just…?
Why not just quantize Equiformer to INT8 or INT4 and skip spikes entirely? This is the most serious threat. The file's evidence for large spiking gains (18–48× on point clouds in 3DSMT, over 360× on graphs in Geometry-Aware Spiking Graph Neural Network) compares against full-precision floating-point networks, not quantized ones. It is also estimated rather than measured. We do not argue around this. Quantized students are a full arm of the study, and the allocation rule is applied to them as per-edge mixed bit width. If uniform INT4 matches the spiking student per joule, we report that spikes add nothing here. The contribution would then reduce to compliance-weighted mixed precision, which is still useful and still new.
Why not just give the spiking network more time steps? More time steps buy precision uniformly, including along stiff bonds where the derivation says it is wasted. The uniform-budget baseline at several total budgets traces this trade-off curve. Our method has to sit above that curve, not merely at a higher point on it.
Why not just train with plain force loss and let the network discover where precision matters? A plain mean-squared force loss is dominated by large forces, which occur along stiff bonds. That is exactly backwards for simulation stability. The loss-only and allocation-only ablations separate the two parts. We expect the weighted loss alone to help modestly, and the allocation to help most when the budget is tight.
Why not use fixed chemical stiffness instead of a learned covariance? This could work nearly as well for small, well-parameterized organic molecules. The fixed-prior ablation settles the question. We expect the learned $\Sigma_i$ to matter more where fixed bond, angle, and torsion models are weak: ISO17 isomers and the larger, flexible MD22 systems, where non-covalent effects omitted by such priors become important.
Proposed Experiments
The primary benchmark is MD17, using its eight molecules. Benzene is included deliberately. It has almost no soft torsional modes, so the method predicts little benefit there, and Fractional Denoising for 3D Molecular Pre-training already identified it as a pathological rigid case. ISO17 and MD22 test generalization to unseen isomers and to larger, flexible systems.
For every model we report four quantities: force MAE, compliance-weighted force error, estimated energy per force evaluation, and spike or operation counts. The decisive metric is simulation stability. We run constant-energy molecular dynamics from held-out conformations and record total-energy drift and the time until bond lengths leave physical ranges.
There are three decisive checks.
- At matched estimated energy, does compliance allocation lengthen stable simulation time over uniform allocation in both the spiking and the INT4 students? If yes for neither, the core claim fails.
- Does the optimal per-edge budget found by a small search track the predicted cube-root law in measured stiffness? A clear mismatch would mean the harmonic argument is the wrong lens.
- Is the gain near zero on benzene and largest on the flexible molecules? That pattern would show the mechanism works through anisotropy rather than through generic regularization.
References
- Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields
- 3DSMT: A Hybrid Spiking Mamba-Transformer for Point Cloud Analysis
- Spiking Diffusion Models
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- Sliced Denoising: A Physics-Informed Molecular Pre-Training Method
- DenoiseVAE: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training
- Deep Spiking Neural Networks for Large Vocabulary Automatic Speech Recognition
- Geometry-Aware Spiking Graph Neural Network
- Fractional Denoising for 3D Molecular Pre-training