Repository navigation
Conversation
…findings
BatchedReplicaExchange._accept_swap implemented only the temperature-ladder
criterion (beta_j - beta_i)(Phi_j - Phi_i), which is identically zero when the
replicas share a temperature. Every mu-ladder swap was therefore accepted with
p = 1, so configurations random-walked freely across the ladder and no replica
sampled its own mu. Replace it with the general form
beta_i Phi_X^(i) + beta_j Phi_Y^(j) - beta_i Phi_Y^(i) - beta_j Phi_X^(j)
where Phi_Z^(k) is a configuration scored with slot k's chemical potentials.
This reduces to the old expression for a shared mu, to beta (mu_i - mu_j)
(N_j - N_i) for a shared temperature, and also covers a joint (T, mu) ladder.
An EMT Ag(111)/Au ladder gives N_Au = [0, 0, 0, 14] with the fix and
[0, 0, 0, 0] without it: unconditional swapping erased the mu dependence
rather than merely degrading the statistics. Existing mu-ladder results are
invalid and need re-running; temperature ladders were unaffected, as was the
MPI ReplicaExchange, which selected _exchange_prob_mu all along.
A degenerate ladder now warns at the end of a run. A 100% swap acceptance sat
unremarked in the log of six consecutive production runs, so the all-accept
and never-accept tallies say so explicitly. Checked on the whole-run tally
rather than live, because a ladder whose replicas have not differentiated yet
legitimately accepts every early swap.
Also fixed:
- PermutationMove with n_swaps > 1 could apply k swaps and then return the
"could not propose" sentinel when a later iteration drew a species absent
from the system. The ensembles read that sentinel as "the atoms were not
touched" and skipped the rollback, leaving the configuration out of step
with E_old. Species counts are swap-invariant, so the usable species are
now resolved once, up front. Both ensembles also restore their snapshot on
the sentinel path rather than trusting the contract.
- Molecules whose centre of mass drifted above a CustomCell's top stopped
being deletion candidates and dropped out of the per-species de Broglie
count, inflating V/((N+1)Lambda^3) into the runaway insertion mode that
docs/gcmc_acceptance_convention.rst describes. Cells grow a second
predicate, is_point_exchangeable, which CustomCell overrides with the same
dropped z upper bound that get_atoms_specie_inside_cell already applied to
single atoms. find_molecules and the displacement one-way-door guard both
use it, so candidacy and displacement regions always agree.
- AlchemiCalculator(energy_only=True) discarded 'forces' from the model
config, which a pre-loaded MACEWrapper shares with every other calculator
built from it, silently disabling FIRE relaxation in an AlchemiFCalculator
depending on construction order. The combination now raises.
- GrandCanonicalEnsemble/CanonicalEnsemble.set_state restored the step count
and exchange statistics from the incoming state. ReplicaExchange passes a
full get_state() dict on every accepted swap, so the two ranks traded their
swap tallies. Only the configuration travels now.
- GrandCanonicalEnsemble.write_outfile returned before reset_counters() when
the outfile was disabled, silently turning interval_ratios() into
total_ratios() for the rest of the run.
- MoveSelector abbreviated move names with a three-character slice, collapsing
MoleculeInsertionMove, MoleculeDeletionMove and MoleculeDisplacementMove to
one indistinguishable 'Mol'. Single-word moves keep their existing labels so
old outfiles stay comparable.
- A missing entry in species_radii raised a bare KeyError from inside the
free-volume sampler; it now names the species.
CI gains the flake8 job CLAUDE.md already documented.
An acceptance-equalized mu ladder for this system needs ~29 rungs (spacing 0.064 eV at the bare end down to 0.020 eV at high coverage, from dmu = 1/sqrt(beta dN/dmu) on the measured isotherm). At ~460 atoms per replica that is ~13k atoms in the relax batch, several times the whole-batch ceiling, so the example could not run a correctly spaced ladder at all. chunk_size ties peak memory to the largest chunk instead of the replica count.
A mis-spaced mu ladder fails silently: no error, no warning, and an output that reads as plausible physics. Two things make it hard to catch, and both are now written down. First, the cumulative acceptance column cannot be used to judge a ladder. Every replica starts from the same configuration, so early swaps are free and inflate the tally permanently. A five-rung CO/CuPd run reported cumulative per-slot acceptances of 28.6 / 19.6 / 5.9 / 2.0 / 0.0 %, which reads as a ladder that merely mixes poorly at one end; in the run's second half three of its four pairs accepted exactly zero swaps and four replicas were independent single-mu chains. The halving test recovers the real rate from the log without extra instrumentation. Second, uniform spacing cannot work across a coverage range at all, because dN/dmu grows with coverage while dmu stays fixed. The fix is dmu = 1/sqrt(beta dN/dmu) from fluctuation-dissipation, with the rung count as the integral of sqrt(beta dN/dmu). For CO on Cu375Pd30 at 400 K over -1.8..-1.0 eV that is 29 rungs rather than 5, spacing 0.064 eV at the bare end down to 0.020 eV at high coverage. Measured second-half acceptance of that ladder: min 18 %, median 40 %, max 65 %, no dead pair, and a monotonic isotherm from 0.1 to 57 CO. The page carries a reference implementation, verified to reproduce that ladder, for whoever lands the mu counterpart of utils.ladder.geometric_temperatures. It also records the calibration caveat (the rule targets the mean-dN exponent, but the realised <min(1, .)> acceptance runs higher, so the rung count is a safe upper bound to trim against) and the two practical consequences of going wide: chunk_size becomes mandatory, and the run gets faster in absolute terms because a narrow ladder leaves the GPU idle.
Minor rather than patch: the release adds a public cell predicate (`is_point_exchangeable`), adds a degenerate-ladder warning, and changes the outfile acceptance-ratio header for molecule moves, which any downstream parser of that column will notice. Also restores the `[1.3.0]` compare link that was omitted when 1.3.0 was cut.
Add the CSI 2026 and AI4AM 2026 abstracts from the group's grand canonical modelling work as evidence of use in the research impact statement, and rebuild paper.pdf.
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.