A potential returning a scalar and an array has not yet produced an interpretable scientific result. You need to know which atoms, geometry, electronic state, units and energy reference those values describe. This guide is for researchers receiving model outputs or preparing a review packet, rather than choosing a model from a performance ranking.
Understand the two projects’ roles
SpookyNet is a reference implementation of a potential architecture with electronic-state inputs and nonlocal contributions, with optional repulsion, electrostatic and dispersion terms. Its original paper explains the motivation. The ability to supply charge or spin-related information does not establish that a particular trained model covers every charge state or reaction.
TorchMD-Net is a framework for training and using neural network potentials. The inspected main README emphasizes TensorNet/TensorNet2 and marks older ET, Transformer and GN architectures deprecated. PyPI metadata reports 3.0.3 and Python >=3.10 on 2026-10-18; main-source behavior must still be checked against the installed release. These projects differ in architecture and software organization; their outputs are not a common benchmark by default.
Establish the input contract
For each configuration preserve atomic numbers, atom order, coordinates, length units, molecule/batch assignment and periodic cell where applicable. Check nonfinite positions, overlapping atoms, missing elements and the model’s permitted boundary conditions before interpreting an energy.
SpookyNet’s core input documentation uses Z, R, per-system total charge Q and magnetic-moment encoding S, plus neighbor information. Its documented convention is singlet S=0, doublet S=1, triplet S=2; do not substitute a quantum spin value or multiplicity without conversion. Its ASE calculator exposes charge and magmom and passes them to the model. A default neutral/singlet setting is not a diagnosis of the molecule’s actual electronic state.
TorchMD-Net’s model source accepts atomic numbers z, positions pos, batch assignments and optional additional inputs. Optional argument names do not establish that every architecture/checkpoint uses them in the same way. Check the selected representation, output module and training configuration.
Read the output as a contract
| Quantity | Interpretation to establish | Common failure |
|---|---|---|
| Scalar per system | Target definition, reduction and energy reference | Calling any scalar a total physical energy |
| Force array | Atom correspondence, negative-gradient convention, energy/length units | Reordered atoms or wrong sign |
| Charge/dipole output | Definition, reference convention and training evidence | Treating model partial charges as unique measured observables |
| Ensemble spread | Which models contribute and how spread was computed | Calling agreement a calibrated error bound |
For an energy potential, forces satisfy F = -∂E/∂R. SpookyNet assumes ångströms and electron volts for its physical terms; its source warns that other units require consistent constant changes. Under that convention forces are eV/Å. Do not merely relabel arrays trained in another convention.
In the inspected TorchMD-Net implementation, derivative=True yields the scalar output and its negative coordinate derivative. The current implementation returns an empty tensor when derivatives are disabled, although its docstring still mentions None; inspect the installed code rather than relying on truthiness or old examples. If the chosen scalar is not an energy, its derivative must not be described as a physical force. Reference offsets, priors and target scaling also need to be traced.
A review exercise without invented numerical values
Create one output packet with: configuration ID, coordinate-file hash, checkpoint hash, code version, target/unit declarations, atom order, total charge/state convention, energy-reference description, raw output shapes and failure status. Start from a public, in-domain geometry with an identifiable reference method. This is a proposed worksheet, not an executed model example; code status is not applicable.
First verify shapes, finite values and atom alignment. Then, for an energy model, propose a finite-difference check on a selected coordinate: compare the reported force component with -[E(R+h)-E(R-h)]/(2h). Choose several displacement sizes and numerical precision deliberately, avoiding pathological geometry or discontinuous neighbor-list changes. Translation and rotation checks should transform vector outputs consistently. These are internal consistency checks: even exact agreement can occur for an inaccurate energy surface.
Evaluate error against the right reference
Compare energy differences only after matching reference method, electronic state, atom composition and energy convention. Atomic reference offsets and total versus per-atom energies can dominate a naive absolute comparison. For forces, specify whether MAE averages Cartesian components or per-atom vector norms; these are different metrics. Group related conformations or trajectories to avoid leakage across evaluation sets.
Include configuration-level errors, subgroups and failed calculations. Examine distorted geometries, unfamiliar charge states and chemistry outside training coverage separately. Good average error on equilibrium data does not demonstrate stable dynamics or reliable bond breaking. Ensemble disagreement may help triage, but low spread is not evidence against a shared systematic error. Experimental observations require their own matching conditions and uncertainty analysis.
Versions, licences and the evidence boundary
The SpookyNet repository and TorchMD-Net code declare MIT licences; checkpoint files and training data still require separate checks. TorchMD-Net’s current README describes compatibility remapping for older TensorNet assets, so preserve the original checkpoint and check the selected release’s migration instructions. This article does not assert binary or numerical compatibility between arbitrary versions.
No model inference, finite-difference calculation, electronic-structure reference or dynamics was executed here. Source review supports the interface statements; the proposed review procedure is editorial guidance. Successful execution, internal derivative consistency, reference accuracy and experimental validation must be reported separately. For execution and broader material-model selection, follow the existing ASE and MLIP Guides rather than treating this article as a ranking.
Related ChemAI Atlas resources and workflows
SpookyNet · TorchMD-Net · ASE
ASE with ML Potentials: Units, Optimization, and Validation · Selecting ML Interatomic Potentials: MACE, SevenNet, CHGNet, and MatGL · Evaluating Molecular Property Prediction on QM9
Sources and evidence status
Sources inspected on 2026-10-18. Primary documentation supports capability/API statements. Audit protocols are editorial proposals; no scientific validation or measured performance is reported.