RELEASE_NOTES.md

Unreleased

Each KinBot and PES run starts its log with the KinBot and RDKit versions.

Conformer restart shortcut for every QC code

A finished conformer search is recognised on restart by the output file copied for the selected conformer, conf/<name>_low.<ext>, together with the last conformer's output. Both were looked up under Gaussian's .log name, so Q-Chem and the Sella-driven codes repeated the whole conformer sweep on every restart. The lookup, the copy and the removal that forces a new search after a lower rotor point now use each code's own output name (.log, .out, _sella.log).

Remove NWChem support

NWChem is not used by anyone. qc: nwchem and the methodclass option are no longer accepted; the supported codes are Gaussian, Q-Chem, ORCA and FairChem.

Remove the NNPES surrogate calculator

The nn_pes calculator depended on a private package that is not available and is superseded by FairChem/UMA. qc: nn_pes and nn_model are no longer accepted.

Remove the barrierless_saddle reaction family

The barrierless_saddle family scanned a user-selected bond, optimized its maximum as a saddle point and wrote a MESS Variational barrier with a CASPT2-oriented Molpro L3. It is removed; hom_sci and VRC-TST remain the barrierless treatments, and the barrierless_saddle* input options are no longer accepted, nor is the never-used variational option.

FairChem compilation option

Add the global fc_compile option, with a default of false. This avoids compilation at the first model evaluation in each short calculation. Set fc_compile: true to enable compilation for all FairChem calculations at L0, L1 and L2, with either a model name or a checkpoint path. Compilation also precomputes UMA's composition-dependent weights to avoid a force-evaluation failure with FairChem 2.14. Other inference settings remain unchanged. Completed results remain usable if this option changes on restart.

Independent calculation levels

• Keep Gaussian AM1 and Q-Chem B3LYP/STO-3G preliminary TS steps when profiles are enabled. The final TS uses L1; scan method options keep their existing meaning. FairChem uses its assigned model throughout.

• Allow executable paths and CPU/GPU devices to change on restart. Record them separately from methods, models, optimizer and frequency settings. Earlier profile records require a fresh directory because search defaults changed.

• Reuse the L1/L2 theory profiles from Hrishikesh Ram's composite branch and extend the same interface to l0_profile. Each level can select its own calculator, method, basis, command or FairChem model/task/device.

• Use assigned profile settings in TS and IRC workers as well as well workers. Read results and Hessians through their recorded level. Global submission resources and L3 remain unchanged.

• Standardize preliminary conformer controls on lowercase l0_*, retaining all previous input aliases, including semi_emp_confomer_threshold. Ordinary inputs without profiles keep their existing calculator and command.

• Changed-profile restarts now request a fresh calculation directory, rather than suggesting removal of the route file while old results remain.

• Record profiles before the first job. Refuse to add profiles to calculations without a matching theory record, or to remove profiles on restart. This prevents results from an old method from being read as results from a new one.

• Remember profile use in .kinbot_run.json, including when profiles and a custom route-file setting are both removed from the input. Existing profiled runs receive this flag when restarted with their original settings.

• Apply the same theory checks to PES before species-input replacement and result reading, including pes ... no-kinbot, without creating calculators.

• Refuse Q-Chem ring preparation before writing or submitting that job; use Gaussian or FairChem at L0 for ring preparation. Q-Chem remains available for ordinary conformer optimization and for L1/L2 calculations.

• Q-Chem profiles require optimizer: "native". Switching a profile to Q-Chem selects that optimizer; Sella/BFGS is rejected before a worker can request an internal Q-Chem optimization in place of a force evaluation.

• Accept profiled_l1_reuse_only: 1 for continuations that must not submit replacement L1 jobs. A missing or incomplete L1 result during a reaction search now fails only that reaction, with its job name in the log. Keep its files and continue other reactions. A missing initial reactant result still stops the run. L0 and L2 remain enabled.

• Refuse unsupported optimizer combinations before submission. ORCA and FairChem require Sella or BFGS. Errors identify settings inherited from the main input when no profile was supplied for the affected level.

• Save each job's level before submission. Refuse rotor results without that record instead of inferring their method from the current high_level flag. The HIR description names the calculator used for the scans, including L2.

• Preserve the optimized product's chemical information when a profiled calculation is saved under its final species name.

• PES and MESS describe the selected L1/L2 settings and retain explicit labels.

• Keep ANL presets and composite-recipe selection out of theory management. Previously written inputs can retain an empty theory_preset field; named presets are rejected. Frequency modes control harmonic calculations only.

Calculation workflow corrections

These corrections also apply to inputs without theory profiles:

• Select preliminary conformers with their L0 energies. Load the complete L1 parent result separately when a failed conformer search needs that result.
• Count actual scheduler jobs when applying queue_job_limit. A limit of one now permits a submission when the queue is empty.
• Use completed calculation status when deciding which intermediate files can be removed. Keep failed calculation files and report the failed step clearly.
• Write Gaussian checkpoint filenames relative to the job directory. Conformer jobs no longer refer to conf/conf/, and checkpoints survive moving the run.
• Read Sella's completion log for Gaussian calculations run through Sella. Keep conf/ and hir/ paths in these logs. Set Guess=Read only when a checkpoint is requested; frequency displacements remove only Read and retain other Gaussian guess options. These corrections include PR #117.
• Record a failed first Gaussian energy call as an error result in the Sella well worker. Use BFGS for order-zero optimizations when requested, while TS optimizations continue to use Sella. A failed TS frequency calculation also writes an error result before the completion stamp.

Job results

A finished job's done stamp is no longer accepted before its result row is in kinbot.db. The job writes its .pkl before the stamp, and over a network file system the driver can see the stamp first. The last row then still belonged to the previous step of the same job, and the next step (for example the final TS optimization of a reaction search) started from that earlier geometry. KinBot now records how many rows a job has at submission and keeps waiting until a newer row exists; the existing 60 s grace period applies.

Identify products after their first optimization

A newly found product can change its connectivity or stereochemistry during its first optimization. KinBot now uses the optimized structure to name the product and saves its complete calculation under that name. This applies to products found through a TS and to homolytic scission products. It does not weaken the discovery IRC check: one direction must still reach the requested reactant. Later well and conformer calculations retain their stereoisomer checks. A product that splits into fragments requires separate fragment calculations; its original frequencies and energy are not copied to them.

First product optimizations use <initial_name>_product. Accepted wells use <optimized_name>_well. If the first calculation from A produces B and the first calculation from B produces C, both results remain available. A restart keeps each reaction's original product calculation. Existing verified wells can supply a first product result without another QC job. As before, reuse of an existing well instead of optimization of each new product geometry means that the order in which jobs finish can affect the selected product. This separation uses calculation format v4; directories made with v3 require a fresh run.

Assign axial stereochemistry from coordinates

KinBot now distinguishes the two orientations of substituted cumulenes. The assignment enters species names, IRC interpretation, reaction-site comparisons, and the existing specified/racemic optical checks. It supports multiple axes and terminal E/Z stereochemistry in longer cumulenes. The temporary graph labels do not change the atoms or masses sent to QC.

Terminal planes within five degrees of a coplanar arrangement receive no axial assignment. No helical identifier is added. The identity and calculation formats have changed; start these calculations in new directories and keep earlier results with their old version.

Single bonds between aromatic rings remain ordinary rotor axes. Opposite twists of a biaryl have the same species identity, so both remain available in HIR scans and MC conformer searches, including in specified mode. The geometric optical check still determines whether the chosen treatment includes the mirror. KinBot does not distinguish fixed atropisomers of biaryls; molecules with a high barrier to this rotation use the same rotor approximation.

The existing absolute phase-space symmetry convention for hom_sci is unchanged. For a chiral parent that forms one chiral fragment, it can give different specified and racemic rate ratios even when the complete mirror reactions should have equal rates. Axial identity support does not correct that separate normalization limit.

Reproducible stereoisomer names from SMILES

SMILES inputs now use RDKit with a fixed embedding seed. Repeated runs from the same SMILES and RDKit version start with the same geometry and stereoisomer name, including when tetrahedral, double-bond or cumulene stereochemistry is unspecified. Explicit tetrahedral and double-bond SMILES assignments are retained; supplied coordinates are unchanged. This selects one starting stereoisomer, not all possible stereoisomers. Use stereochemical SMILES or coordinates to request a particular stereoisomer. If the first embedding fails, KinBot retries once with random starting coordinates and the same fixed seed. If both attempts fail, KinBot asks for coordinates. The ETKDG method is unchanged. When MMFF94 cannot refine a structure, KinBot uses UFF if its parameters are available. A warning reports missing force-field parameters or incomplete refinement; QC can still use the generated coordinates. Disconnected fragments are placed apart so the generated coordinates do not create false bonds between them. OpenBabel remains in use for other helpers.

Stereochemistry and symmetry

Keep configured stereoisomers separate throughout a calculation. KinBot keeps chemid as its connectivity identifier. A stereochemical identifier distinguishes configured tetrahedral centers, double bonds, and axes. Ordinary species keep their existing calculation names. Configured species use different names for QC jobs, saved results, reused objects, and direct/PES output. The same comparison checks IRC endpoints and fixed configurations during conformer searches.

Configured calculation names use -s followed by 16 hexadecimal characters. The full stereoisomer identifier remains in saved records and is checked before reuse. Names do not depend on optical_population: a racemate must still remain separate from its diastereomers. The shorter names use a new calculation format; calculations with the earlier 64-character suffix require new directories. If an input SMILES leaves stereochemistry unspecified, the log reports the assignment in the generated geometry at INFO level. This note also covers partly specified SMILES. It is not emitted for supplied coordinates.

RDKit >= 2025.9.3 is required. KinBot selects its stereo-perception settings explicitly and records these settings and the RDKit version in the log. SMILES input uses RDKit to generate the starting geometry. An unsupported initial reactant stops before QC. An unsupported discovered reaction is omitted with a warning that the network is incomplete. Its calculation files remain available; unrelated reactions can continue.

The four stereochemical H-transfer example files are valid JSON templates. Replace their quoted SET_... placeholders with local QC and queue settings before running them. Replace the CPU count and job limit strings with integer values. The HIR templates use matching placeholders for method and high_level_method, and for basis and high_level_basis; keep these pairs equal when replacing the placeholders for an L1-only HIR calculation.

Ordinary PAHs and biaryls are no longer rejected by their number of aromatic rings. Scope checks use an aromatic graph copy so equivalent phenyl arms do not appear different in one bond drawing. Aromatic radical resonance forms are not mistaken for cumulenes. Virtual reaction-site labels are used only for comparisons; they do not label the physical molecule. The geometric optical calculation still determines whether a harmonic, HIR or MC model includes its mirror. No helical stereoisomer identifier is added: selection of one fixed helical enantiomer and separation of helical diastereomers remain unsupported.

Start older calculations in new directories. This release cannot resume calculations from older KinBot versions. It does not convert old job names, directories, or result formats. Each new calculation has a .kinbot_run.json format record. A restart requires that record and the same RDKit version and settings. KinBot checks this before it changes old files. Save an environment specification, such as a conda export or pinned Python requirements, with each project so that this runtime can be restored.

New feature: select one stereoisomer or a racemate

Treating a chiral well as one specified stereoisomer is new. In earlier KinBot versions, a chiral well identified by the graph-based chirality rule was treated as a racemate: the code assigned nopt = 2 and wrote SymmetryFactor = sigma_ext / 2 in MESS. There was no option to restrict that well to one enantiomer. The old rule detected atoms with four different bonded groups; it did not detect all forms of geometric chirality.

The new input option has two values:

• "optical_population": "specified" is the default. It includes the stereoisomer assigned from the input geometry and excludes its opposite enantiomer.
• "optical_population": "racemic" includes that stereoisomer and its whole-molecule mirror. This restores the earlier counting of both enantiomers for those chiral wells. It does not include other diastereomers. For example, an RR input includes RR and SS, not RS and SR.

Existing inputs change behavior when this option is omitted. KinBot reads the stereochemistry from the geometry; the input does not need an R/S label or an explicit stereo option. Every run with the option omitted uses specified. For a chiral well with external rotational symmetry number sigma_ext = 1, represented by one structure, the previous SymmetryFactor 0.5 becomes SymmetryFactor 1.0. To retain the previous factor for this well, add "optical_population": "racemic". SymmetryFactor is a divisor: 0.5 includes twice the statistical contribution of 1.0 for the same structure.

Example: sec-butylperoxy, CH3CH2CH(OO•)CH3. Its tetrahedral stereocentre has H, methyl, ethyl, and peroxy groups. Use charge: 0, mult: 2, rotor_scan: 0, and multi_conf_tst: 0 to compare one harmonic structure. For the supplied enantiomer, sigma_ext = 1, and MESS receives these factors:

Treatment	Enantiomers included	SymmetryFactor
Earlier KinBot versions	Both	0.5
New default, or explicit specified	The input enantiomer	1.0
Explicit racemic	Both	0.5

This change does not imply a factor-of-two change in every rate coefficient: the TS contribution also depends on which reflected reaction its reactants and products permit. stereo_reference preserves the requested stereoisomer in generated PES inputs.

Conformational mirrors of achiral molecules are also counted. For example, a single gauche ethanol conformer has a distinct mirror although ethanol has no fixed stereocentre. With rotor_scan: 0 and multi_conf_tst: 0, earlier versions wrote SymmetryFactor 1.0; the new treatment writes 0.5 when the missing mirror is established. Both conformers belong to the same stereoisomer, so this also applies in specified mode. A HIR model that already includes both mirrors, or an MC model with both mirrors explicitly present, receives no extra factor of two. Thus racemic restores the earlier choice of enantiomers, but it does not undo other corrections to optical or rotational counting.

Reaction pathways and statistical counting

Find rings without generating every open atom path. Ring detection now enumerates closed cycles directly. It keeps the earlier ordered ring list, including larger perimeters of fused rings. This reduces the cost for PAHs without changing the resonance-structure search.

Keep different stereochemical reaction pathways. Virtual substitution distinguishes homotopic, enantiotopic, and diastereotopic sites in the related reaction motifs. It does not change the atoms sent to QC. The checks include heavy atoms and reactions with several reacting sites. Forming and breaking bonds distinguish reacting atoms from spectator atoms. Global atom-equivalence groups remain unchanged. When equivalent searches merge, KinBot prefers the earlier graph search's atom selection if it passes the reaction-family checks. For example, propane keeps r12_insertion_R_7_2_3 instead of renaming it to r12_insertion_R_7_2_1. Distinct stereochemical selections still receive separate searches. Direct MESS and PES output keep different stereochemical pathways, also when lowestpath is selected. Repeated observations within one pathway still use the existing lowest-barrier selection. A MESS Union adds the different pathway contributions without requiring MC-TST.

Use the symmetry and properties of each conformer. Earlier MC-TST output could write symmetry-equivalent conformer copies as separate RRHO blocks and overcount them. The old moments-of-inertia filter was ineffective because all() returned a boolean instead of testing each moment ratio. Duplicate checks now compare geometries under equivalent atom permutations. Each selected conformer keeps its geometry, electronic energy, ZPE, frequencies, Hessian when available, calculation source, and status together. MC population filtering uses each conformer's rotational and optical weights. MESS then uses those same weights. Comparisons allow equivalent atom permutations and proper rotations, so the same conformer is not counted twice. Duplicate geometry comparisons use an E + ZPE window of 0.5 kcal/mol. This allows more numerical variation than the earlier 0.2 kcal/mol window while avoiding expensive comparisons between well-separated energies. Explicit mirror comparisons retain their separate treatment without this energy gate. Explicit mirror pairs receive no additional optical multiplier. Conflicting numerical observations produce warnings. They do not require the whole calculation to stop. MC-TST continues to disable HIR scans. An L1 or L2 MC result with missing or non-finite energy or ZPE now fails the affected optimization with a warning. Other reaction calculations continue. The code does not substitute zero for these missing conformer properties. Internal record-association errors still raise an error.

Keep the graph-based rotational (external) symmetry rules with limited corrections. Fixed stereochemistry prevents exchanges between incompatible configurations. Pyramidal XY3 has a threefold rotational contribution. Planar XY3 has sixfold. The pyramidal rule and optical factor assume a harmonic single-well inversion mode; a future explicit umbrella model that covers both minima must account for their symmetry and mirror contribution without counting them twice.

A departing atom is distinguished from the spectator atoms when rotor (internal) symmetry is calculated.

Apply one optical-counting method to harmonic, HIR, and MC models. The method compares local molecular parts and accounts for the active rotor bonds to assign optical symmetry factors. The local geometric RMSD tolerance to identify optical isomers is 0.1 angstrom. Existing HIR scans are used to establish mirror coverage. If a HIR model already includes both mirrors, KinBot adds no second factor of two. If an allowed mirror is clearly absent, its contribution is included.

For unresolved comparisons within the allowed optical population, KinBot uses the selected calculation's Hessian to estimate the stable-mode harmonic energy at an aligned mirror midpoint. Each MC conformer uses its own Hessian. The optical factor is one for an estimate at or below 4 kcal/mol, and two above that value. This is an approximation. The estimate is not a calculated inversion barrier. If the necessary data remain unavailable, KinBot uses factor one with a warning. optical_factor_assumptions permits an explicit factor of one or two for a named structure, with a necessary written reason.

Keep usable rotor scans when another rotor fails. Shared records contain units, calculation sources, raw and projected frequencies, rotor axes, symmetry numbers, scan angles, energies, and point statuses. Optimization replaces scans whose geometry or rotor definition no longer matches the selected structure. It retains compatible scans, including usable partial scans, and keeps the old calculation files. A failed rotor retains its harmonic motion; the other rotors remain in the model. Projection starts from the full Hessian of the selected calculation. If that Hessian cannot be used safely, KinBot restores all harmonic frequencies and omits HIR with a warning. MESS writing can check saved data and repeat this projection, but cannot submit QC or wait for scans. Missing scan coordinates alone keep the usable rotor potential and give a warned optical factor of one. A scan that changes to an excluded stereoisomer or reaction pathway is unusable; an optical factor of one cannot correct such a potential. The existing rotor-zero energy check still disables all rotors if their common optimized reference is inconsistent. The initial L1 result is loaded completely when no later optimization replaces it. Reused product calculations keep the accepted structure and its properties.

Separate disconnected well networks. KinBot follows reactions between bound wells from each requested reactant. Separated-product channels are exits. A shared product does not connect two otherwise disconnected well networks. Each requested disconnected network gets its own MESS input and output. For example, this prevents specified R-butanol from acquiring an unrelated S-butanol network through a common product.

New feature: choose whether MESS includes product complexes. me_skip_vdW: 0 is the default. It includes accepted product complexes and their barrierless exits in both single-well and PES MESS output. me_skip_vdW: 1 omits these wells and exits and connects each inner TS to the separated products. Complex searches and optimizations still run. Inner-TS Eckart depths use the selected complex energy with either setting. Both writers use the existing PES approximation of one lowest-energy complex for each set of product stereoisomers. MC-TST retains its selected-structure tunneling reference. Omitting a complex removes its stabilization and explicit capture/redissociation competition, so rates can change. PES complex selection no longer depends on discovery order. Previously, the first complex could keep its own energy reference even when a later complex for the same products had a lower energy. All complex names for that product set now resolve to the lowest-energy complex. correct_submerged now uses connected bound-well ground energies only. A TS below separated fragments is not raised to their energy.

Complete MESS jobs before reporting their result. Local, Slurm, and PBS execution waits for the requested calculations to finish. Earlier versions skipped MESS execution when queuing: "local" was selected. With run_me: 1, a local run now starts MESS and waits for the result. With run_me: 0, KinBot writes inputs and scripts without starting MESS. Local execution, queue scripts, and manual scripts now use mess_command; earlier versions ignored this option and used mess. The option specifies an executable and optional arguments. For example, use "mess_command": "/opt/MESS/bin/mess". Quote a path or argument that contains spaces. Shell variables and command substitutions are not expanded. Relative paths are resolved from me/. Input-only writing needs no local MESS executable; queued jobs find the executable in the compute node's environment. A local run reports an error if it cannot start the configured command. KinBot checks the exit result and newly written output. Solver failure is reported separately from successful reaction generation. Inputs without an accepted reaction network remain available for inspection, but are not run. KinBot writes its reaction summary and PESViewer input before starting MESS. For queued MESS jobs, it allows up to 60 seconds after the job leaves the queue for the exit marker and new output to appear. A recorded solver failure still fails immediately. Scheduler queries have a 30-second timeout and allow three consecutive failures before reporting an error; a query failure does not mean that the job has finished. Initial input-reference and saved-result conflicts reported by StereoRoutingError are logged and exit with failure, without an internal traceback. Other internal errors are not suppressed.

Output formats and reaction-discovery cost. Summary files include a # kinbot_stereopath {json} comment before each successful calculated saddle, including ordinary pathways. Homolytic-scission placeholders can lack this comment. External summary readers must allow comment lines; the PES reader uses this metadata to keep different pathways separate. MESS microcanonical files use <job-stem>.micro, such as me/mess_0000.micro, instead of the shared micro.out. Uncertainty samples and disconnected networks have separate filenames. me/mess_networks.json lists the current MESS jobs. For ordinary QC calculations with high_level: 0, the MESS header now reports the L1 method and basis instead of unused L2 settings. This corrects the header description; it does not change the calculation level. The database includes one stereochemistry/<job> input-reference row per well optimization job checked for reuse. It records the input geometry, full stereoisomer identity, charge, and spin multiplicity. It is not another QC result, and job-status checks do not repeatedly add it. Database readers must distinguish these reference rows from calculation results. Reaction discovery performs additional stereochemical graph comparisons. A ten-molecule check with fixed coordinates measured about 2.2 times the enumeration time of master. This timing excludes imports, characterization, initial stereoisomer-reference assignment, and QC. It is not a measurement of the total calculation time.

Narrow bug fixes and reaction examples. H2 elimination now rejects longer ring walks when an equivalent shorter search was removed as a duplicate. Distinct stereochemical paths remain separate in PESViewer output, also when their barriers differ by less than 1 kcal/mol. An exact stereoisomer entry takes priority over a connectivity entry in vrc_tst_scan and vrc_tst_noscan, including an empty list. The same rule applies before VRC calculations and when PES writes rotdPy input. The initial species reconstruction keeps its checked frequencies. Input-only MESS writing converts collision-energy units in the same way as normal writing. Three-fragment product names agree with each other through PES assembly. Long configured names use bounded file names for MESS intermediates. Four peroxy H-transfer inputs compare diastereotopic pathways with an achiral control, using HIR or MC-RRHO. Their collision parameters are illustrative. Tests include saved methanol, peroxy, pyramidal, and loose-TS structures.

The new stereochemical rate counting targets MESS. MESMER receives the same species names but does not implement the new pathway and optical sums.

Wait for delayed calculation results. After a submitted Slurm or PBS job leaves the queue, KinBot allows up to 60 seconds for its complete result to become available. This prevents an early failure or a duplicate submission when result files appear late. A completed calculation error is still returned without this extra wait.

Native Q-Chem constraints are written correctly (#67). The six Q-Chem job templates imported ASE's stock QChem calculator, which knows nothing about KinBot's addsec keyword and wrote it into $rem as ADDSEC $OPT ..., which Q-Chem rejects ("Illegal rem input in read_rem"). KinBot's own calculator, which writes the constraints as a separate $opt ... $end block, has existed since 2023 and was already used by every Gaussian and Sella template. The Q-Chem templates now use it too. This affected every native Q-Chem transition-state search and hindered-rotor scan. Testing against a real Q-Chem also turned up four further defects, fixed in the same change: bond and angle constraints were computed but never written (only torsions reached the $opt block); releasing a constraint indexed the geometry with one-based atom numbers and raised IndexError; the Total energy = summary line printed by Q-Chem 6.2 was not recognised, so converged jobs reported no energy; and an empty OPT keyword made Q-Chem frequency jobs run another optimisation. Validated on Q-Chem 5.4.2 and 6.2.1.

PES post-processing writes pesviewer.inp first and no longer draws its own graph (#82). The pyvis-based "interactive graph" duplicated what PESViewer does from pesviewer.inp, and it was the only post-processing step with a third-party dependency that could fail or hang for reasons unrelated to the kinetics; a PES whose KinBot runs had all finished could end without a combined pesviewer.inp. The graph is removed and pyvis dropped from the plot extra. pesviewer.inp is now written immediately after the energies are assembled, before the rotdPy and MESS inputs, and each post-processing stage announces itself in pes.log.

Near-linear molecules keep 3N-5 vibrations, and MESS agrees. get_frequencies decided how many external rotations to project out by an absolute cutoff of 1e-5 on the mass-weighted rotation vectors. Any optimised geometry that is linear only up to an optimiser residual (a bend of 0.01 degrees is enough) exceeded it, so three rotations were projected and one component of the degenerate bend was lost: 3N-6 frequencies instead of 3N-5. With Gaussian or Q-Chem this reached MESS only through the rotor-projected frequency set; with FairChem or Sella, where all frequencies come from this routine, it affected every near-linear species (CO2, HCN, C2H2, ...). Linearity is now decided by three gates: all bond angles within 2 degrees of 180; the curvature of the Hessian along the rotation about the molecular axis is a vibration (above 50 cm-1), not a free rotation, which distinguishes a linear molecule left slightly bent by the optimiser from a genuinely bent minimum; and that curvature exceeds three times the residual of the other rigid-body motions. Stored geometries are never modified. A species judged linear is written to the MESS input as an exactly linear rotor, with a comment, so that MESS's own moment-of-inertia test (I_min/I_mid < 1e-5) reaches the same conclusion; otherwise MESS would treat a 179-degree CO2 as a non-linear top with a tiny third moment, a factor of ~3 in its rotational partition function. The reverse mismatch is handled too: a genuinely bent minimum so close to linear that MESS's test would call it linear gets its three rotational constants written explicitly, so MESS keeps the non-linear rotor that the 3N-6 frequencies assume. Both decisions are logged. The same routine now uses the symmetric eigensolver, so degenerate modes can no longer come back as complex numbers.

Generated job scripts no longer depend on numpy reprs (#84). Element symbols and coordinates were pasted into the ASE/Molpro job templates via the repr of numpy types, which under numpy >= 2 reads np.str_('C') and np.float64(1.06). Scripts from the five templates that do not import numpy (Q-Chem TS search, IRC and HIR; NWChem constrained TS search; Molpro TS search) failed to run. Values are now converted to builtin Python types before formatting.

2.4.1

Bugfix release. Two groups of changes alter numerical output for inputs that ran under 2.4.0 and earlier; they are listed first.

Results change for unchanged inputs

Hindered-rotor projection (#99). The internal-rotation vectors that are projected out of the Hessian were built from mass-weighted atomic coordinates instead of mass-weighting the rotational displacement. For any atom whose mass differs from the rotor-axis atoms this skews the vector, so the projected frequencies handed to MESS (reduced_freqs) were wrong for every rotor_scan = 1 run since KinBot 2.0. On butane the corrected projection reproduces the full spectrum minus the three torsions to within 9 cm-1, where the old one deviated by up to 590 cm-1. Because the error largely cancels between a transition state and its reactant, the effect on rate coefficients is typically 10-20% at combustion temperatures; quantities without a cancelling partner (equilibrium constants, standalone entropies and heat capacities, master-equation densities of states) see the full single-species error, up to a factor of 1.5-2 for molecules with several rotors at high temperature. The regression reference in tests/frequencies.py now comes from a finite rigid rotation of the molecular fragment rather than from the code itself.

Multi-conformer TST (#100).

• Direct (non-PES) runs never resolved the per-member energy offsets, which were written only as comments; every Union member sat at the parent energy. Offsets are now written as numbers on both routes.
• Member energies are referenced to the accepted parent structure instead of to the lowest member, so a conformer that becomes lowest only at L2 is placed at its true energy rather than displacing the whole Union.
• Each saddle member gets its own imaginary frequency and shifted Eckart depths; members whose L2 optimisation failed are excluded; a single surviving member is written as a one-member Union with its own properties.
• Population screening in find_unique passes the complete mode list to ASE's IdealGasThermo (API from ASE 3.23, already implied by the ASE 3.26 requirement). Previously all imaginary modes were dropped before the call, which raises under current ASE for any transition state.
• Conformer ground energies are initialised from E + ZPE at L1, so L1-only MC-TST members are no longer equally weighted.

Free rotors (#100). The MESS free-rotor block now carries the rotor Symmetry number, as the hindered-rotor block already did. A methyl free rotor was previously overcounted threefold.

Fixed

Selected-structure consistency (#100).

• With conformer search at L1, the rotor projection used the Hessian of the starting structure with the geometry of the selected conformer. The selected calculation's geometry, energy, ZPE, frequencies and Hessian are now loaded together.
• Every poll of the optimisation re-ran the conformer selection and reset the species geometry to the L1 conformer after the L2 geometry had been loaded. The selection now runs once.
• When no Hessian is stored for the selected structure (Gaussian conformer jobs, native Q-Chem L1), a frequency-only calculation at the fixed geometry recovers it. If recovery fails, KinBot warns and continues with harmonic frequencies and no hindered rotors.
• The accepted result is published under the conventional job name (<name>_well, conf/<name>_low or <name>_well_high), so restarts and PES post-processing read the same structure the optimisation accepted. Replaced output files are archived, not deleted.

Hindered-rotor restarts (#100). With high_level = 0, a lower point found during a scan is now re-optimised at L1 before the scans restart, instead of being taken as a stationary point directly from the constrained geometry. Restarts archive the superseded scan results through the database rather than deleting log files. qc_opt_ts honours its high_level argument.

Miscellaneous. Q-Chem L1 jobs print the Hessian; formchk is invoked through subprocess with error checking instead of a silent os.system.

Known issues

• Rotors whose far side has only multiply bonded neighbours (nitro, nitroso, isocyanate groups) are not detected, because dihedral detection requires a single-bonded terminal atom.
• The semi-empirical conformer pre-search is not yet restricted to wells.

Contributors

Luka Dockx, Judit Zádor.