Conversation
- examples/lammps/ALIGNN_FF_LAMMPS_Colab.ipynb: end-to-end Colab notebook (install ALIGNN, build LAMMPS with pair_alignn, download default mps model, relax Si, run NVE + melt-quench, plot results). Linked in README. - scripts/torch/lammps_package/: drop-in ML-ALIGNN LAMMPS package mirroring the LAMMPS source tree (src/ML-ALIGNN, cmake module, doc page, example), with install steps and an upstream-PR checklist.
The notebook called get_model.py / build_si.py / *.in which were untracked, so a fresh clone (and the Colab run) couldn't find them. Commit the example source/input files (not the generated .pt/.data/.lammpstrj artifacts), and fix the notebook to clone the develop branch where these examples live and to read the atom count from si.data instead of hardcoding it.
…ers/FrechetCellFilter)
get_prediction now uses the DGL-free pure-torch path when pure_torch=True or when DGL is not installed. Adds load_pure_torch_model, get_figshare_model_pure, and get_prediction_pure, plus a --pure_torch CLI flag. Pure models are pulled from all_models_alignn_atomwise.json (mps, formation_energy_peratom, mbj_bandgap, bulk_modulus_kv, ...), built with build_pure_torch_graph, and run via ALIGNNAtomWisePure(_Smooth).
pair_style now gets <cutoff> <max_neighbors> (both read from the model config if unset), matching what the model was trained with. Adds --fail-force/--fail-energy thresholds and a non-zero exit on gross pair_alignn-vs-Python mismatch so it can be used as a pre-MD gate.
pair_alignn.cpp computed periodic cell offsets with diagonal box lengths only (round(disp/h_i)), which is wrong for non-orthogonal boxes. A 3x3x3 supercell of JVASP-1002 (the 2-atom primitive FCC cell) is triclinic, so shifts were wrong on ~90% of ghost edges (edge-vector errors up to ~23 A), producing garbage forces and exploding MD (T -> 1e6 K). Replace with an exact triangular back-substitution against the LAMMPS lower-triangular lattice rows; reduces to diagonal rounding for orthogonal boxes. Applied to both the dev pair style and the ML-ALIGNN package copy. Also: build_si.py now uses the conventional (cubic) cell so its supercell stays orthogonal -> works correctly even with an already-compiled binary. Verified: ASE NVE on the cubic cell conserves energy (~0.15 meV/atom over 250 fs); triclinic shift recovery is exact (2000/2000) vs 196/2000 before.
The ASE AlignnAtomwiseCalculator builds a DGL neighbor graph, which differs (~0.1 eV/A) from the pure-torch k-nearest graph that pair_alignn implements and the mps model was trained with (neighbor_strategy=pure_torch). Comparing against it produced false gate failures. Default reference is now 'pure' (build_pure_torch_graph + forward_tensors_z), the apples-to-apples baseline; --reference ase keeps the old behavior.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.