Libra v5.10.0: Bohmian dynamics, PyTorch DVR, FISH, revised atomistic workflows

Human-added content

Major additions:

• Bohmian dynamics module in PyTorch - uses multivariate Cauchy-Lorentz trajectory basis functions to represent the probability density, automatic differentiation to compute derivatives needed to construct quantum potential and differentiate it to obtain quantum forces, uses partial vectorization for efficiency; this release comes with the revision of the manuscript "Stable Direct Dynamics with Quantum Potential: Lorentzian Trajectory Basis Function is All You Need" that describes the methodology and shows the method's performance;
• Highly-efficient DVR solvers via PyTorch - implement exact numerical solution of TD-SE for adiabatic and non-adiabatic dynamics. Way faster than the previous C++ implementation. Very convenient to use as well.
• NA-MD recipes - a collection of standardized recipes defining various quantum-classical methodologies now comes with Libra. This should minimize the inconsistencies of defining such recipes by different users.
• FISH (fully-integrated surface hopping) - this method was already available in the previous version, but now it is finalized with few minor fixes. The relevant publication is now available: Han, D.; Shakiba, M.; Akimov, A. V. "Fully-Integrated Surface Hopping as Quantum Decoherence Correction in Nonadiabatic Dynamics" J. Phys. Chem. Lett. 2025, 16, 28, 7168–7176
• LDR (local diabatic representation) approach for exact quantum dynamics on a grid - a new module written in PyTorch is added and is nearly finalized, the corresponding article in preparation.
• Interface to ODIN code for correct calculations of time-overlaps with DFTB+ - this is now added, the corresponding DFTB+/Libra interface and workflows are revised. In general, this helps obtain significantly more reasonable results than before, but care should be exercised when using SK files - only correctly-generated ones can be used.
• Revised atomistic workflows - in particular, the calculations of the overlaps and time-overlaps of Slater determinants is now corrected, the mapping functions are revised, the phase factors previously given in the code are dropped. This does affect the calculations of TD-DFT overlaps and time-overlaps, and helps minimize the presence of "sudden jumps" in population dynamics. At the same time, our tests show that previous workflow still yielded close numerical results in many cases

Minor changes and additions:

• improved documentation, especially regarding the installation procedures,
• fixed bugs in Ferretti's model
• added several model Hamiltonians written with PyTorch

Auto-generated content

What's Changed

• Debug the Ferretti's model by @DaehoHan in #254
• Added the LDR solver with PyTorch by @DaehoHan in #255
• Fix coefficient initialization in LDR by @DaehoHan in #256

Full Changelog: v5.9.0...v5.10.0