Add cumulative strain guard to FixSymmetry to prevent phase transitions#445
Open
janosh wants to merge 2 commits intoTorchSim:mainfrom
Open
Add cumulative strain guard to FixSymmetry to prevent phase transitions#445janosh wants to merge 2 commits intoTorchSim:mainfrom
janosh wants to merge 2 commits intoTorchSim:mainfrom
Conversation
janosh
force-pushed
the
fix-symmetry-cumulative-strain
branch
3 times, most recently
from
1aa6711 to
89496fe
Compare
The per-step deformation check (max_delta <= 0.25) is necessary but not sufficient: many small steps can accumulate into a Bravais lattice change (e.g. hexagonal → tetragonal) that breaks the symmetry constraint. This was observed in production on a Sc-Nb-B structure with c/a ≈ 70 (c = 209.6 Å). During NequIP OAM-L relaxation, the c-axis collapsed to 6.2 Å through individually small steps, losing all hexagonal rotations. The fix stores the initial refined cell as a reference in from_state() and checks the total deformation from that reference in adjust_cell(). When cumulative strain exceeds max_cumulative_strain (default 0.5), the cell update is scaled down proportionally to stay within the envelope. The per-step deformation check is also softened from a hard RuntimeError to a clamp (scaling the step to max 0.25), since the cumulative guard is now the real safety net. This fixes test failures where rotated structures with noisy forces occasionally produced single steps just over the 0.25 threshold (e.g. 0.2581). - Add reference_cells and max_cumulative_strain to FixSymmetry.__init__ - Store refined cells in from_state() for cumulative tracking - Clamp per-step deformation > 0.25 instead of raising RuntimeError - Clamp cumulative strain in adjust_cell() when limit exceeded - Propagate reference_cells through reindex/merge/select - Add regression test with direct adjust_cell loop verifying clamping - Update test_large_deformation to verify clamp behavior (not raise) Co-authored-by: Cursor <cursoragent@cursor.com>
janosh
force-pushed
the
fix-symmetry-cumulative-strain
branch
from
89496fe to
5f78e2d
Compare
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
2 participants
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.
Summary
Builds on #438 (FixSymmetry constraint). Adds a cumulative strain guard that prevents phase transitions in structures with extreme aspect ratios.
Problem
The existing per-step deformation check (
max_delta <= 0.25) catches large single-step cell changes but not slow cumulative drift across many steps.Fix
reference_cellsinfrom_state()adjust_cell(), compute cumulative strain from the reference cell (not just the per-step delta)max_cumulative_strain(default 0.5), scale the cell update proportionally to stay within the strain envelopereference_cellsthroughreindex(),merge(),select_constraint(), andselect_sub_constraint()The 0.5 default allows substantial legitimate relaxation (50% strain from the initial cell) while blocking Bravais lattice changes.
Test
Adds a regression test with an elongated hexagonal cell (c/a ≈ 70, SG #191 P6/mmm) under heavy compression (0.7x). Verifies:
Note
This is complementary to #443 which fixes L-BFGS history contamination from unsymmetrized forces. That PR addresses the force-direction drift; this PR addresses the cell-deformation drift. Both failure modes can cause symmetry loss on pathological structures.