The Forth Bridge. Widely believed to be painted continuously since 1890.

This list is kept against the development log. Items done since August 2026 stay here for a while with the month they were done, so that anyone who knew the old list can see what moved. Items in hand carry a name.

Done since August 2026

(By Mike Towler working much harder than usual, except where noted)

New methods and measurements

• DMC that stops itself: when the error bar reaches a requested target, or at the smallest error bar a given time budget can buy.
• popbias: the population-control bias of a finished DMC run, and the energy corrected for it, with error bars.
• EQUIL_DIAGNOSTICS: after a DMC run, where two standard rules would have ended equilibration, and how far the answer would then have moved.
• LIMDMC 7 and 8: the reweighting of Anderson, Per and Umrigar, which its authors found reduces time-step errors at large time steps.
• VMC_STEP_PROFILE: smaller VMC steps near all-electron nuclei, making VMC of all-electron atoms and small molecules heavier than helium two to five times cheaper (Konjkov/MDT).
• Densities with error bars: a zero-variance estimator (ZV_DENSITY) for molecules and crystals, error bars on the ordinary density, and expval2cube to write them as cube files for VESTA and other viewers.
• The exchange-correlation hole and the on-top pair density of a molecule, with error bars (HOLD_PARTICLE), and hole2html to drag the electron through the hole.
• The localisation tensor of a correlated state, which tells an insulator from a metal.
• Reptation Monte Carlo repaired: pure expectation values with error bars, every process’s reptile combined, and reptiles that no longer freeze near a node.
• Light nuclei, such as protons and muons, treated as quantum particles alongside the electrons in DMC.
• The probability of finding a given number of electrons in a region of space (DOMAIN_PROB).
• DMC estimates of the energy’s derivatives with respect to determinant coefficients (DMC_NODE_DERIV), a first estimate of the nodal gradient.
• Nodes made visible: nodeview draws a wave function’s nodal surface in a web page, nodecells counts its nodal cells and tests Ceperley’s tiling property, and wfnscope lets you drag the electrons and watch the node, the local energy and the conditional density move.
• pilotdrive and RUNTYPE pilot_eval: de Broglie-Bohm trajectories in superpositions of CASINO wave functions, and the relaxation of their ensembles to quantum equilibrium.

Wave-function optimisation

• Energy minimisation now tries the full step it computes, which its search had replaced by a meaningless one, usually refused.
• Analytical Jastrow derivatives corrected: several had been wrong, and cycles with fixed cutoffs had scaled them by unrelated factors for over a decade.
• Variance and MAD minimisation no longer stop early on false or singular convergence after a rejected step.
• Analytical derivatives for backflow parameters and cutoffs, determinant coefficients and f-term cutoffs: optimising backflow in finite systems is about six to seven times faster.
• Reweighted variance minimisation much less prone to run away (VM_VARW_MAX), and a warning when an optimisation cycle leaves the energy or variance far above the best so far.
• CHECK_ADERIVS checks the derivatives the optimisers actually use.
• Optimisation of Slater-type orbitals repaired: in some inputs it changed nothing, and it tied unrelated coefficients together and left redundant rotations free.
• Geminals: open shells supported and optimisable, constraints applied, manual and help current.

Backflow and Jastrow factors

• The three-electron backflow term omega keeps the wave function antisymmetric again, is sampled exactly, and has all its spin triplets optimised.
• The backflow Phi term’s cusp conditions, applied to the wrong coefficients since version 2.0 whenever its sets had different orders, are now right.
• Backflow works with periodic all-electron Gaussian orbitals and the cusp correction, where energies could be wrong by tens of hartrees a cell.
• A parameters.casl Jastrow term that excludes a nucleus no longer counts it: DMC_METHOD 2 energies of water had been 0.52 Ha too low.

Getting the right answer

• All-electron DMC_METHOD 2 branches with its own factor, not the T-move one it had used since 2018.
• Parallel DMC_METHOD 2 runs no longer use stale or uninitialised Jastrow values on walkers moved between processes, a bias of about 0.1 mHa per atom in silicon.
• DMC walkers are reweighted correctly after a change of wave function, with space warping in DMC molecular dynamics, and at a change of twist.
• Gaussian shells are screened on what they can contribute, removing errors of several mHa per cell in crystals such as nickel oxide.
• Cusp corrections of the virtual orbitals of multideterminant expansions are right; wrong ones could change the fixed-node energy markedly.
• Compressed multideterminant expansions with zero coefficients, as at the end of large CIPSI expansions, no longer change the wave function.
• Gaussian f and g functions: wrong g-function gradients under screening and wrong f-function values under periodic backflow corrected.
• VMC jumps between ions no longer bias the energy.
• NaN energies stop the run in optimised builds, where the tests had been compiled away and NaN results reported.
• POPSTATS no longer prints a failed estimate as a number, and warns when a run is too short for it.
• Slater-type orbitals with n = 5 normalised correctly; n up to 10 supported.
• Forces: planarity test and per-atom charges fixed, periodic forces refused, all-electron forces flagged as unreliable.
• Positrons: each ion type has its own positron pseudopotential, and a positron in a molecule sees the ions correctly.
• Ewald sums correct on nearly symmetric cells in 2D and 3D; the 1D Coulomb sum accurate to 10^-10 hartree per pair.
• A core-polarisation potential’s l-dependent term no longer cut off at the pseudopotential’s radius.
• Momentum densities and the condensate fraction weighted correctly in DMC.

Speed

• Faster Ewald sums: the reciprocal-space part of the 3D electron-electron sum now grows linearly with the number of particles rather than as its square, and slabs share the work between ions in a plane.
• Gaussian orbitals evaluated faster: DMC 17 to 33 per cent faster for a cluster of 64 water molecules.
• Slater-type orbitals screened (STOTOL) and evaluated faster: VMC of C₃₂H₆₆ 2.5 times as fast from the screening, and 1.7 times faster again from the evaluation.
• Faster T-moves, multideterminant DMC and Jastrow F term.
• DMC can share walkers by measured core speed, for desktops with fast and slow cores (REDIST_SPEED_WEIGHT).

Other codes

• Cartesian d, f and g Gaussian shells read natively, as in 6-31G* and other Pople basis sets; converters say what they can produce.
• CASINO checks at start-up that a molecule’s Gaussian orbitals are orthonormal, catching converters that got the basis wrong.
• MOLDEN route completed and tested for PSI4, DALTON, Molpro, ORCA, NWChem, OpenMolcas and TURBOMOLE.
• New converters: trexio2qmc for TREXIO files, Quantum Package’s CIPSI expansions included; qbox2pwfn and casino2qso for QBOX; crysgen23 for CRYSTAL23.
• GAMESS converter remade and tested with GAMESS 2024, f and g functions included.
• Q-Chem through molden2qmc: its MOLDEN files in spherical or Cartesian d, f and g functions, its restricted open-shell orbitals, and its own output read directly.
• AMS/ADF: the adf2stowf converter distributed with CASINO.
• CASTEP: runcastep runs CASTEP and converts its orbitals in one command, and twistav_castep resumes an interrupted calculation.
• Quantum ESPRESSO 7.6: runpwscf and runqmcmd read its files.
• CRYSTAL: runcrystal brought up to CRYSTAL23 with exit statuses; g functions from CRYSTAL in molecules; opt_crystal repaired, with new -kappa and -stall options; a manual appendix on choosing and optimising CRYSTAL basis sets for QMC.
• 2DHF numerical orbitals for diatomic molecules usable again, from x2dhf 3.0.
• ABINIT: how to run it for CASINO and give it a CASINO pseudopotential.
• MCEXX: wrote completely new converter (old one had disappeared)
• nwchem_pp_convert correct for ccECP and for NWChem’s own libraries.
• mpc.data generated automatically when a crystal’s MPC interaction needs it; WRITE_GEOMETRY writes the atoms for viewing.
• A Tersoff bond-order potential among the manual interactions (NDD).

Running CASINO

• CASINO, runqmc and autotest exit with a failure status when a run fails, and parallel runs stop rather than hang when an error reaches only some processes.
• Wave-function optimisation and DMC in a single run.
• Many malformed or unsupported inputs, over-long input lines among them, now refused with a message instead of being silently misread.
• runqmc no longer appends a new run to an earlier run’s output (--keep-old keeps the earlier files).
• Builds with Intel ifx, an optional MPI-module build with argument checking, and OpenMP shared-memory groups fixed.
• Regression suite extended: RMC, geminals, backflow, analytical derivatives, optimisation with fixed cutoffs and more.
• reblock: unconverged quantities and short-run POPSTATS warnings.
• Pictures from a run: WALKER_FILM films the walkers of a DMC or VMC run, frame by frame, with the numbers that tie each walker to its parent, and the new utility walkerfilm makes of the film a trajectory for OVITO or Blender or a web page that plays it, in place of the old movie, which is withdrawn with movie2avi

Documentation

• Most of the manual rewritten for new users, with a new opening chapter on QMC and CASINO and a new chapter on reptation Monte Carlo.
• Every input keyword’s help rewritten; the manual’s keyword reference is now generated from the code, so the two agree.
• Correlated sampling with vmc_sampling 'stdtest' documented.

LOUIS

• LOUIS 1.3: a user manual, an ensemble mode with Nelson’s stochastic dynamics, and corrected 3D densities.

 

What we would like done next, by subject, the larger jobs first in each. An item carries the name of whoever is on it, so that two people do not do the same thing; an item without a name is free. If you want one, say so on the forum or write to Mike Towler, and it is yours. The developers’ page says how a change reaches the code, and the development log shows what became of the last list.

Open: new methods and measurements

Things CASINO cannot yet measure, or cannot yet do, that people ask for.

• Correlated sampling in DMC. VMC has it (vmc_sampling stdtest); DMC does not, and it is what an adsorption energy, a reaction barrier or the small displacements of a geometry optimisation need, since two independent runs resolve a difference of a millihartree only at a cost that grows as the square of the ratio of the local energy’s spread to that difference.
• Forces for plane-wave and blip orbitals and in periodic systems, where they are refused today: the derivative of the ion-ion Ewald sum, the change of the orbitals with the nuclei (by space warping, or from the generating code), core-polarisation potentials and f and g pseudopotential channels. Then a geometry optimiser that knows its forces carry error bars.
• Excited states on the same footing as the ground state: state-specific optimisation (variance matching, and Zhao and Neuscamman’s variational principle for excited states) beside the promotion method, and the optical gaps of solids with twist averaging, as a documented workflow.
• Light nuclei in the rest of CASINO: DMC molecular dynamics, reptation, and a hydrogen atom moving among fixed nuclei in a metal.
• Nuclear quantum effects by averaging: electronic energies averaged over nuclear configurations drawn from a vibrational wave function or a path-integral run, with correlated sampling to take the differences cheaply.
• Neural-network trial wave functions: a plug-in evaluator, value, gradient and Laplacian at a configuration, called at every move, through which DMC runs on a FermiNet or Psiformer wave function or on a learned Jastrow factor. DMC on such functions exists elsewhere; the point is CASINO’s T-moves, twist averaging and expectation values on them.
• Spin-orbit coupling: two-component spinor determinants from relativistic CRYSTAL and DIRAC calculations, with spin-orbit pseudopotentials, in VMC and DMC. The complex and spinor Gaussian evaluators are done (v3.1.391 to 402, October 2026); the readers of the two codes’ output, the spin-orbit potentials with their projectors and the sampling of the spin coordinate are the work left. Mike Towler, with the CRYSTAL developers.
• The probability of finding n electrons in a region of space (DOMAIN_PROB): Voronoi cells, spheres, half-spaces or a Bader basin read from a cube file; P(n), its spin-resolved form, the mean, variance and spin of each region and every pair’s covariance, with error bars, in VMC, DMC and RMC. It extends POPULATION, which gives the mean count alone, and it is the Scemama probability function the old list asked for, with the charges, spins, moments and fluctuations. Mike Towler; under review.
• Bader charges and the topology of a QMC density with error bars, from the cube files CASINO now writes, as a utility rather than a recipe.
• A uniform electric field as an input keyword for finite systems, in place of a hand-made external potential; with the dipole moment CASINO already measures, polarisabilities follow from one field. Mike Towler; under review.
• Electrons in separated subsystems as distinct particles, for bilayers and double quantum wells where antisymmetry between the layers costs much and buys nothing.
• A criterion that reports an unbound trial wave function for a small complex (a positronic or excitonic one, say) instead of a slowly drifting energy.
• Embedding: a QMC region inside a classical or DFT environment, beginning from the external potentials CASINO already reads and going on to an embedding potential derived from a DFT density.

Open: wave functions

Jastrow factors, backflow, pairing and determinants: the forms CASINO can optimise and run.

• The general Jastrow factor and general backflow as the documented defaults, their cost against the standard forms measured on real systems before the switch.
• Open-shell geminals with backflow, and with a complex wave function, which cannot be combined today; then Pfaffians (PSI_S pfaffian is refused today), which hold geminals and determinants as special cases and allow triplet pairing.
• Hybrid orbitals, Gaussians plus blips, for a molecule on a surface or a defect in a crystal, where one basis serves neither part well.
• A library of starting Jastrow terms by element, the electron-nucleus and electron-electron-nucleus terms that carry from one system to another, so that an optimisation starts near its answer.
• The plane-wave p term of the Jastrow factor re-represented in blips, so that it costs what a blip orbital costs
• DMC energies with geminal wave functions and backflow fixed (hadn;’t worked for many years).

Open: wave-function optimisation

Getting the parameters right, and knowing when they are.

• An audit of energy minimisation’s convergence, with the gates a run must pass before its parameters are trusted: the condition of its matrices, the quality of its line-search fit and the stability of its lowest eigenvector, reported in the output rather than left to the user’s eye.
• Optimisers for very many parameters at once: a descent method, or the sample-space forms of the linear method and stochastic reconfiguration whose cost grows with the number of samples rather than of parameters, for the general Jastrow, backflow and the coefficients of a selected-CI expansion of a hundred thousand determinants together.
• An optimisation that stops when it has converged, as a DMC run now stops at its target error bar, instead of running the number of cycles it was given.
• Analytical derivatives for what still uses finite differences, the general Jastrow’s cutoffs and basis among them.

Open: getting the right answer

Correctness, and the checks that prove it.

• All-electron forces: an estimator with an honest error bar, or a refusal. Today they print, flagged as unreliable, because the bare estimator’s variance is infinite; the renormalised estimator of Assaraf and Caffarel and the filtered one of Chiesa, Ceperley and Zhang are the known remedies.
• Post-weighted expectation values in parallel DMC (TPDMC above zero) use the walkers of the first process alone, so the densities, pair-correlation functions, structure factors, density matrices and the rest come out with a fraction of the statistics and error bars to match; the energy is unaffected. Found in October 2026. Mike Towler.
• The DMC time step: Anderson and Umrigar’s 2021 form of the T-move (their reweighting is in, as LIMDMC 7 and 8), measured on pseudopotential systems; and a time-step extrapolation check in the analysis tools, since two extrapolations of the same data can disagree by millihartrees.
• A continued run that refuses what has changed: every accumulation carried between runs written with the layout that made it, so that an older version’s accumulation is never silently mixed with a new one. The node-derivative estimator continued with a changed set of eps values is the case in hand.
• A reference set that catches a wrong answer and not only a changed one: the regression suite compares a build with the outputs an earlier build wrote. Add benchmark energies with published values, atoms, molecules and solids, the hydrogen chain of Motta and co-workers (2017) among them, rerun at every release.
• The suite reading every file a run writes, not the out file alone: the history files, config.out and expval.data are written by every test and compared by none.
• A test deck for every mode: 0D to 3D, every orbital type, VMC, DMC and RMC, DMC molecular dynamics included, and a census that shows what no deck reaches (mpc.data generated in the run, STOP_METHOD, reptation at the bare nucleus of an atom heavier than helium).
• Every expectation value checked under DMC weighting, twist averaging and VMC’s alternative sampling against an independent estimate: under the alternative sampling every expectation value is unweighted today, and several weightings were found wrong in 2026 by reading rather than by any test.

Open: speed and scale

A run of measured speed-ups has shipped since August; these are the levers left, each to ship with its measured timing.

• CASINO on GPUs beyond the present OpenACC build: the Gaussian and blip evaluators, the Jastrow factor and the Ewald sums on the device, measured on the standard test systems against the processor at equal cost.
• Large multideterminant expansions from selected CI: the work done once per determinant shared over distinct spin strings, for real wave functions first, projected at several times faster at a hundred thousand determinants.
• Parallelisation over twist angles, each twist its own DMC population, so that a twist-averaged run fills a whole machine at once and the twists finish together; and no re-equilibration after a twist change that leaves the occupancy unchanged.
• Scaling measured again: MDT ran CASINO on many hundreds of thousands of cores in 2011 and nothing has been measured above a few hundred ranks since. A scaling study on a national machine with today’s code, and whatever it finds, fixed. (Having once had simultaneous accounts on four of the top-ten most powerful machines in the world, MDT has been brought low and is currently trying to prove that Vallico Academy is a real institute so he can apply for time on more than 64 cores. Embarrassing. If anyone can help let him know.)
• Fault tolerance for very large runs: a run that survives the loss of a node and continues from its last configurations.
• Configuration files written and read by MPI-IO, so that a checkpoint of a million walkers is not a serial bottleneck.
• More of the large arrays in shared memory within a node, as the orbital coefficients already are.
• OpenMP that pays: on a 64-core workstation two threads a rank are slower than two ranks, so the threading buys memory and nothing else.
• Reptation rebuilds the Slater matrix about thirteen times a move: 44 per cent of an all-electron carbon run. Mike Towler.
• The pairing of G vectors with their negatives in the expectation-value set-up is quadratic with no early exit: a density on a long polymer never finishes starting.
• The periodic Gaussian contraction is called five times per derivative evaluation per k point where one call taking all five basis vectors would do. Mike Towler.
• The backflow Phi term forms every basis product in its innermost loop, about twice the arithmetic it needs; the same in the Omega gradient and Hessian.
• Numerical orbitals are all evaluated and then masked, where only the masked ones are needed; and the periodic cusp correction zeroes large arrays on every call.
• FFT grids for mpc.data that are not powers of two, and FFTW as an alternative library.
• The cost per move at a thousand electrons profiled and attacked, as was done this year at a hundred, the localised orbitals and the Ewald sums first.

Open: external codes

Routes for trial wave functions into CASINO; the interfaces page says what works today. A new Gaussian-basis route is usually a MOLDEN or TREXIO file, which CASINO’s converters read already; otherwise a route needs a writer of pwfn.data, gwfn.data, bwfn.data or stowfn.data, and the interface guide says how. We help.

• WIEN2k and FHI-aims: all-electron orbitals from an LAPW code and a numeric atom-centred code, the first all-electron periodic routes besides CRYSTAL. WIEN2k needs an LAPW orbital type, radial functions times spherical harmonics inside the spheres and plane waves or blips between them, which would serve exciting and elk as well. Mike Towler.
• A numerical atom-centred orbital type in CASINO, which the FHI-aims route needs and which would open SIESTA, OpenMX and DMol3 too.
• Converters for VASP (with norm-conserving potentials) and OpenMX. Aleksey Baglov, exploring.
• MCEXX: a writer of CASINO’s files inside MCEXX (October 2026), to be had from the CASINO developers on request until MCEXX itself includes it; the route’s validation through DMC is under way, then its manual entry and its line on the interfaces page. Mike Towler.
• CFOUR: the corrected reorderdf.f that its MOLDEN output needs is distributed with CASINO; the route is untested.
• GAUSSIAN: the converter is twenty-five years old, spherical functions only, and unmaintained; a current route is wanted. Mike Towler.
• TREXIO as the common currency: CASINO reads orbitals and determinants from it today; next, pseudopotentials and Jastrow parameters, and writing CASINO’s own trial function to it, so that one wave function can go to CASINO, QMCPACK, TurboRVB, CHAMP or an AFQMC code and the methods be compared on equal terms.
• CP2K, Octopus and BigDFT: untried. CP2K writes MOLDEN files for molecules, so molden2qmc first; the real-space and wavelet codes need a conversion of their orbitals to blips.
• Blip converters inside CASTEP and ABINIT, as Quantum ESPRESSO has, so that the DFT code writes blips itself and no pwfn.data of tens of gigabytes is written and read.
• Consistent treatment of unoccupied orbitals across the plane-wave interfaces, so that an excitation means the same thing whatever code made the orbitals.
• Point-by-point verification carried to every route: orbitals and densities compared with the generator’s own at sample points, as the MOLDEN routes were in 2026, so that every entry on the interfaces page has a measured verdict.
• Gaussian basis sets built for QMC rather than borrowed from quantum chemistry: the successor to billy and opt_crystal (not the one embedded in CRYSTAL23!), and the study of what an optimised basis does to Hartree-Fock, DFT and DMC energies. Mike Towler.
• Pseudopotentials: spin-orbit potentials for the spinor route; every potential of the library in every code’s format from the one converter CASINO ships; and the library’s element pages brought up to date with the ccECP set added in October 2026.

Open: running CASINO

Installing it, driving it, and reading what it says.

• Defaults that let a DFT user get a trustworthy DMC number: a Jastrow factor chosen and optimised, the time step and population set from the system, the time-step and finite-size extrapolations done, and runqmc’s checks carrying the manual’s advice. The “make the answer better” button, built one default at a time.
• Installation: an installer that works without a terminal, a built tree that can be moved, a download link that names its version, a “check my install” that runs a few systems against reference energies, and known-bad compiler versions refused at compile time.
• A Python layer: an ASE calculator that writes the input, runs runqmc and reads the out file, and a workflow for the time-step and finite-size extrapolations that carries every error bar through to the number quoted.
• In the output, a cost line, the product of run time and energy variance, which says what a smaller error bar will cost before the run is extended; and a line on the health of the branching, the spread of the weights and the rate of catastrophes, beyond what POPSTATS reports.
• Builds and architecture files for today’s machines kept current, ARM (Grace, Apple silicon), AMD, Intel ifx and NVHPC among them, with the suite run on each at every release.

Open: documentation

For users first, then for developers.

• Complete rewrite of the manual, every example checked against the text, and the reference generated from the code kept so. Mike Towler.
• A first week with CASINO: ten worked examples from H₂ to a crystal, each with its reference numbers and the time it takes on a laptop.
• A guide for DFT users: what DMC does for you, what it costs, how to read an error bar, and when a fixed-node answer can be trusted.
• The programming guide brought up to the code as it is, with a map of the source: which routine does what, and where a new expectation value, orbital type or interaction goes.
• The FAQ remade from the forum’s questions since 2012.
• Every example in the distribution running with its reference output, and the gaps in what the examples cover filled.

Open: LOUIS

The quantum trajectory code distributed with CASINO.

• The velocity loop: timing a shorter-chain table, and a discrepancy in one dimension. Mike Towler.
• LOUIS and CASINO’s pilot-wave driver reading and writing one trajectory format, so that a film of a two-electron atom and of an ensemble of four thousand atoms come from the same tools.

Open: infelicities

Small, well-defined faults: a first patch for a new developer.

• POPULATION stops when expval.data exists without its block; and it goes on accumulating through the second, electron-fixed run of a PAIR_CORR calculation, so that its weights double.
• A DMC catastrophe in one twist of a twist-averaged run zeroes every expectation value’s sums, the earlier twists’ included.
• plot_expval’s histogram error bars on reptation data are too small by a factor of two to ten, and it loops on an invalid answer at the end of its input.
• The nickel oxide 3x3x3 example shipped with CASINO has a pathological Jastrow factor: its final variance evaluation gives a total energy of -483190 au against about -1500, and its header says 4x4x4.
• Energy minimisation’s eigenvector: the debug build picks a spurious one of an ill-conditioned problem when the manipulation constant is near 10¹⁰, where the optimised build does not.
• The scaled-sum norm routines in CASINO’s own lib (dnrm2 and its kin) report finite norms for a NaN beside a tiny or huge element in an optimised build; building them with -fno-finite-math-only cures it and moves no number.
• The start-up file the installer writes, .bashrc.casino, uses HOSTNAME, DOMAIN, ARCH, KERNEL, DISTRIBUTION and OS as scratch variables and unsets them all, the user’s HOSTNAME with them.
• The FIXMEs in the source (grep finds them), and the dead code: the unused special functions in lib/specfun, and runcrystal’s CRYSTAL95 and 98 branch, which it has refused since v3.1.329.

From the last list: the Scemama probability function is DOMAIN_PROB above; the DMC-MD Green’s function is part of the light-nuclei item; twistav_castep now resumes interrupted runs and its error traps go with it; complex orbitals with a Gaussian basis are done (v3.1.391 to 402, October 2026) and belong in the done list above; symmetry operations in xwfn.data, high-throughput QMC (now the defaults item) and charges, spins and moments (now DOMAIN_PROB) are dropped as separate items; the Blue Gene/Q items and asynchronous I/O were dropped earlier as obsolete.