Skip to main content
Back to timeline
arXivSource publication:

Excited states of strongly correlated open-shell systems skip the intervening ladder: spin-adapted neural backflow reaches chemical accuracy for O2's singlet

Synopsis

Excited-state calculations for strongly correlated systems previously had to optimize every state below the target in energy order, and near-degenerate states of other spins made that costly. Spin-adapted neural-network backflow (SA-NNBF) confines the wavefunction to a prescribed spin sector and extends it to multiple sets of spatial orbitals to restore universality. For the lowest singlet of O2, single-state variational Monte Carlo reaches a 1.46 mHa error at D=1, inside the 1.59 mHa chemical-accuracy threshold; combined with natural excited states, the four lowest triplets of O2 and the three lowest singlets of [Fe2S2(SCH3)4]2– also reach chemical accuracy.

Source-provided article image: Spin-Adapted Neural-Network Backflow Wavefunctions for Excited States of Strongly Correlated Open-Shell Systems
Figure 1 · arXiv · Page 2

Interpretation

Excited-state calculations can now be carried out directly inside a prescribed spin sector, without optimizing every intervening state of every spin below the target. Earlier multi-state frameworks such as natural excited states had no total-spin control, so reaching the first excited singlet of [Fe2S2(SCH3)4]2– required first optimizing the 28 lower-energy states in the Ms=0 spectrum. In the (20o,30e) active space of [Fe2S2(SCH3)4]2–, spin adaptation reduces the same target to 2 states, against a full-spectrum count of 29.

The construction has a neural network output only spatial-orbital transformation matrices, which are combined with a fixed spin eigenfunction into spin-orbital coefficient matrices, then antisymmetrized as determinants over occupied spin orbitals and summed. Configurations that differ only in spin assignment share the same spatial matrix, and the wavefunction is an eigenfunction of total spin and spin projection by construction, so no spin penalty term is needed. On the triplet ground state of O2 (6-31G basis, (18o,16e) active space) at comparable parameter counts, the preferred number of determinant sets D for SA-NNBF is typically at least 4, versus 1 or 2 for vanilla neural backflow.

Expanding the spatial orbitals from one set to D sets makes SA-NNBF with a fixed spin function able to represent exactly any state in the chosen orbital space and spin sector once D is sufficiently large. The single-set version lacks this property: all configurations with the same spatial occupation share one matrix, so the number of independent equations can exceed the unknowns that matrix supplies. On the lowest state of each of six spin sectors of [2Fe-2S], SA-NNBF(1,128,1) missed chemical accuracy for half of them, while raising D to 2 brought errors consistently within 1 kcal/mol, comparable to or better than vanilla NNBF(1,128,1) at similar parameter counts.

Perspective

The demonstrated scope is limited to active-space models of O2 (6-31G basis, (18o,16e) active space) and [Fe2S2(SCH3)4]2– ((20o,30e) active space), with accuracy measured as energy error relative to FCI and chemical accuracy taken as 1.59 mHa or 1 kcal/mol.

The universality result requires a sufficiently large number of determinant sets D; in practice D is finite, and the choice of spin function may affect accuracy and optimization difficulty. The authors also note that more compact spin-function decompositions and a multi-state formulation scaling more favorably with the number of states remain open.