How to understand this CHANGELOG (and FHI-aims):

AI guidelines

The FHI-aims AI guidelines were added to the source code. You can find them in the file AIMS_AI_GUIDELINES.md.

They explain how AI tools may be used when working with or developing FHI-aims, how to protect the FHI-aims source code, and how to ensure that AI-assisted work complies with the FHI-aims license. They also address human responsibility for reviewing AI-generated suggestions and respecting third-party licenses when contributing code or other content.

The AI guidelines were discussed and decided by the aims-coordinators. This is the first version of the AI guidelines and we want to review and improve them over time. Thus, we appreciate any feedback you can give. Also, if something remains unclear or you have specific questions, please feel free to ask.

Changes since release version 260331

25 different authors among more than 180 merged MRs.

---

Relativistic Q4C Forces and Stress (Wentao Zhang)

• Added support for forces for relativistic Q4C calculations at the LDA and GGA level for periodic systems. (!2130, !2180). This also enabless structure relaxations with Q4C (!2303).
• Bugfix: update the frozen W matrices of q4c method in SCF iteration (!2282)

---

Magnetic Anisotropy Energy (Xingtao He)

• The Magnetic Anisotropy Energy (MAE) can be computed from the following set of keywords in control.in:

    include_spin_orbit
    calculate_mae .true.
    mae_axis 1 1 0 0
    mae_axis 2 0 0 1

This will compute the MAE based on the SOC corrected eigenvalues as the difference between the x axis and z axis.

• Note that the MAE needs a dense k-grid to get converged.

---

Grand-Canonical DFT (Brad Ayers)

• Constant-potential (grand-canonical) SCF: the electron number floats while the chemical potential is held at a target electrode potential; the headline energy becomes the grand potential Ω = E\(_{KS}\) - μN, reported alongside the canonical energy components at every iteration and in the final output (!2267).
• New keywords gcdft, gcdft_target_potential, gcdft_reference_potential, gcdft_mu_tolerance, gcdft_mu_damping, gcdft_allow_jellium, gcdft_omit_mu_n_term.
• Net electrode charge is compensated by the MPB solver or an Environ embedding; convergence additionally requires |eF - μ| below gcdft_mu_tolerance.
• FHI-aims manual section “Grand-Canonical DFT” added, including basis-set guidance for charged electrodes.

---

Skala Functional (James Green, Sebastian Kokott, Konstantin Lion, Andrei Sobolev)

• Support for the machine-learned XC functional Skala (G. Luise et al, arXiv:2506.14665) has been added (!2230). Skala functional evaluationscan be run on either CPU and CUDA.
• To compile FHI-aims with Skala support, enable USE_SKALA and provide a LibTorch installation through LIBTORCH_PATH:

    set(USE_SKALA ON CACHE BOOL "")
    set(LIBTORCH_PATH "<path>" CACHE STRING "")

For installation with CUDA please see manual, or ftorch CI pipeline (cmake/toolchains/ftorch.intel.cmake)

• Skala can be selected in control.in: xc skala 1.1. The corresponding model weights must be available. They can either be placed in the working directory under the predefined model name or specified explicitly using skala_model_path. Model weights for Skala 1.0 and 1.1, as well as several baseline functionals, are supported. They are distributed under MIT license with FHI-aims source code in external_libraries/Skala/models directory
• Skala is currently an experimental feature:
  • Restricted to molecular systems; periodic boundary conditions are not supported.
  • Forces and stress are not implemented.
  • Results from Skala-1.1 are also dependent on the grid partition scheme, so the same partition_type should be used consistently when comparing calculations (inherent to Skala itself).
  • Precompiled LibTorch distributions may cause conflicts with system mathematical libraries, in particular Intel MKL.
• When Skala is requested, FHI-aims prints an experimental-feature warning and stops unless override_warning_skala is specified in control.in.
• More info: see part of manual chapter 3.3 on Skala functional.

---

Symmetry-broken initial density (James Green, Sebastian Kokott)

• Added a framework for initializing from symmetry-broken free atoms (initial_density orbital_occupations). In addition, initial orbital occupations can now be specified through initial_occupation keyword. (!2293). The framework is especially helpful for converging free atoms faster.
• initial_density orbital_occupations is now default for free atoms; the old default can be switched on under initial_density spherical_atoms
• initial_density spherical_atoms remains the default for multi-atomic systems.

---

Updates for LibRPA and Periodic GW (Minye Zhang)

• Updates to the library LibRPA (!2206):
  • spinor wave-function parsing
  • two-component periodic GW calculations for k-grid/band quasiparticle energies
  • improved band interpolation to conserve degeneracy at small k-grid
  • improved QP equation solver
• Updates to the interface (!2221 !2258)
• Added batch LVL coefficient generation for LibRPA parsing and symmetry-operation export. (!2263 !2233)
• Fixed LibRPA band output for hybrid functionals (!2279 )
• Rank-1 update method for periodic GW head correction (!2311)
• Documentation updates (!2162)

---

Berry Phase Polarization Improvements (Elia Stocco, James Green)

• Substantially reduced memory use in the ScaLAPACK polarization implementation. use_local_index and load_balancing are supported for parallel polarization calculations. (!2147)
• Added a parallel LAPACK implementation of polarization (!2153)
• Related bugfixes and improvements (!2138 !2227 !2295 !2204)

---

Friction and Superconductivity Improvements (Connor Box)

• Output improvements:
  • Added HDF5 output for frequency-dependent electronic friction. (!2181)
  • Friction output now includes separate spin-channel results for collinear-spin calculations and can write EPCs for occupied-to-occupied and unoccupied-to-unoccupied transitions. (!2173 !2140)
  • Restored friction elsi epc write, fixed the associated finite-difference matrix handling, and added state indices to EPC output. (!2172 !2118)
  • Added atom-resolved phonon DOS and atom-resolved Eliashberg-function output for superconductivity. (!2141)
• Updated friction acoustic-sum-rule handling and enabled the superconductivity “grid” mode. (!2171)
• Fixed the ScaLAPACK analytical-friction tensor for clusters and enabled ScaLAPACK for the corresponding mode-2 calculations. (!2201)
• A number of changes to improve the numerical stability of the superconductivity code, in order to be more aligned with established workflows (!2239):
  • A new keyword for superconductivity, alpha2f_frequency_cutoff, excludes modes with frequencies below this value from construction of the Eliashberg function.
  • Eliashberg function construction switched back to the “mode” style from pre-May 2026
  • Lambda and derived quantities are now calculated directly from the modes - not from integrating the Eliashberg function

---

Transport Module Improvements (Xingtao He)

• Added support for use_local_index in transport calculations and updated the corresponding control-file handling. The implementation also addresses reference-potential alignment by retaining the sparse overlap information needed to rebuild the overlap matrix. (!2207 !2214)
• Added manual selection of atoms used for transport potential-reference alignment via the per-atom pot_alignment keyword. (!2149)
• Fixed transport energy-step handling. (!2274)

---

Meta-GGA minimal basis from sratom atomic solver (Sebastian Kokott)

• Enable LibXC meta-GGA functionals for the sratom atomic solver (!2242 !2259).
• This feature can be enabled by setting use_meta_gga_sratom .true. in control.in.

---

Forces and Stress Improvements

• Reordered density-matrix based force and stress term contractions. This gives speedups of up to a factor of two for the force/stress evaluation. The computation of kinetic density and density gradients during density update was also optimized. (Sebastian Kokott - !2300)
• Enabling tuned path from Florian Merz for calculating Hartree potential terms for stress and force correction term, speeding up forces and stress evaluations (!2305)
• Fixed the long-standing spin-collinear meta-GGA stress bug caused by passing xc_tau_deriv with an incorrect array layout. (James Green - !2156)
• Analytical meta-GGA stress calculations received additional bug fixes and checks (Mariana Rossi, James Green - !2216 !2112)

---

RPA Forces for Spin-polarized Molecules (Muhammad Naeem Tahir)

• Enabling RPA forces for spin-polarized molecules (however, only integer occupations supported) (!2039 !2043 !2047 !2049 !2053 !2054 !2050 !2056 !2059 !2061 !2062 !2063 !2064 !2066 !2068 !2069 !2071 !2072 !2073 !2076 !2079 !2081 !2085 !2088 !2089)

---

Sternheimer RPA for Finite Systems (Hao Peng)

• Added an RI-based Sternheimer RPA correlation-energy path for finite atoms and diatomic molecules, activated with use_rpa_stern_diatom, with documented control options and restrictions. (!2157)
• Added documentation and regression testing for the RI-Sternheimer RPA implementation, including explicit Sternheimer grid controls and diagnostics for the dense LAPACK fallback solver. (!2165)
• Added an iterative Sternheimer RPA path for finite single atoms and diatomics using the existing real-space Hamiltonian construction and ARPACK response eigenvalue calculations. (!2196)

---

Bethe-Salpeter/GW

• Added dynamical (frequency-dependent) BSE functionality. (Songrui Liu - !2299)
• Optimized memory use in full-relativistic (2c) GW. (David Lara - !2294)
• Fixed the correct building of xc_matr_2c for Spin Collinear calculation in Full Relativistic GW (David Lara - !2254)
• Fixed bugs in importing wavefunctions/spinors for 1c/2c GW calculations (Qing-Long Liu - !2256)

---

ELPA 2026 and ELSI Updates (James Green)

• ELPA version 2026.02.002 is now shipped with FHI-aims via ELSI, and can be built internally in FHI-aims (see MR !2298 and ELSI MR !394).
• This version should be more performant on GPU, and adds AMD/HIP GPU support.
• The manual section “H.3 Building ELPA 2026” has been added, detailing how one can build FHI-aims with this version of ELPA, and elpa2026.intel.cmake and elpa2026.cuda.intel.cmake toolchains have been added to cmake/toolchains for intel LLVM CPU and GPU (CUDA/NVIDIA) builds, respectively.

---

Other Improvements

• Use meta-GGA for XC evaluation during innitialization step (PBE remains fallback for hybrid meta-GGAs). (James Green - !2148 !2154)
• Added xc r2scan keyword as shortcut for the r2SCAN libxc functional, avoiding the need for override_warning_libxc because the functional combination is explicitly supported and tested. (Andrei Sobolev - !2304)
• Implemented the periodic version of Δ-SCF through deltascf_basis (Matthias Kahk - !2187 !2215)
• Added JSON-formatted dipole communication to i-PI. (Mariana Rossi - !2137)
• Continued development of domain-decomposition RI (renamed from “gloRI” to “ddmRI”), including an MPI shared-memory redistribution method and an AMD-GPU linking fix. (Moritz Leucke - !2128)
• Added RI-LVL support for RT-TDDFT. (Chris Shepard - !2245)
• Removed unnecessary _potential routines from get_ri_coeffs. (Alan Lewis - !2281)
• Released XDM(Z) from experimental status (Kyle Bryenton - !2151)
• Updated XDM(BJ/Z) damping parameters following changes to the lightdenser basis and added a corresponding manual note. (Kyle Bryenton - !2193)
• Added a check preventing use of the original TS correction for systems containing alkali species unless explicitly overridden with override_vdw_correction_hirshfeld_alkali. (Andrei Sobolev - !2160)
• Output the electronic group velocity (k-derivative of the electronic eigenstates) on a k-grid for transport calculations, for example interfaces with TransOpt3.0 to compute Seebeck coefficients, electrical conductivities, and electronic thermal conductivities (Jingkai Quan - !2243)
• Added core and CPU-domain information to the output and reorganized prepare_system checks. A runtime check for a known Intel MPI Bcast/Allreduce problem on AMD hardware was added (Andrei Sobolev - !2296)
• Updated aimsChain to be a Git subproject, analogous to ELSI, making it easier to use and update. (Andrei Sobolev - !2186)
• The final energy block now repeats the spin-collinear moment summary, including the integrated and absolute spin-density differences. (Xingtao He - !2203)

---

Bugfixes

• Fixed a Hermiticity bug in spin-orbit coupling calculations. (Qing-Long Liu - !2277)
• Some bugs were fixed in the on-site and mulliken subroutines of the +U code (Lydia Fichte - !2228)
• Improved periodic hybrid-functional band-structure interpolation in the case of small structures and loose k-grids; added additional Fourier interpolation checks (Minye Zhang, James Green, Andrei Sobolev - !2270 !2219 !2191)
• Fixed an array-layout issue in the spin-polarized libxc call from sratom. (Sebastian Kokott - !2302)
• Fixed a bug in Ca minimal+s species defaults. (Thomas Purcell - !2241)
• Fixed additional whitespace in MD_light output. (Konstantin Lion - !2155)