Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set
summary
The gist
Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set is introduced to construct small virtual spaces out of large basis sets for…
In short
The SVD-TC method compresses large virtual spaces for transcorrelated calculations using singular value decomposition (SVD). It projects canonical orbitals from a large basis onto a smaller one, creating efficient virtual spaces. This allows high accuracy from big basis sets while avoiding the complexity of traditional reference correction methods.
Key concepts
- Singular Value Decomposition (SVD)
- SVD is a mathematical technique used to decompose a matrix into three simpler matrices. In this context, it's used to project large virtual molecular orbitals onto a smaller set of atomic orbitals, effectively compressing the virtual space while retaining important information.
- Transcorrelated (TC) Calculations
- These are advanced quantum chemistry calculations that go beyond standard Hartree-Fock methods by including electron correlation effects. The SVD-TC scheme applies this correlation correction efficiently by working within a compressed virtual orbital subspace instead of using composite reference corrections.
- Virtual Space Compression
- This involves taking a large set of virtual molecular orbitals from a big basis and projecting them onto a much smaller, selected set of atomic orbitals. This reduces the computational cost significantly because the expensive correlation steps only operate on this much smaller, compressed space.
Terminology used across episodes
This episode discusses
- Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set · Paper Radio
- An Additive Reference Correction Scheme for the Transcorrelated Method
The paper
Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set · Read on arXiv
Max Planck Institute for Solid State Research · University of Cambridge
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Compression of virtual spaces in transcorrelated methods via singular value decomposition".
Mira: Here's the gist of the paper. Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set is introduced to construct small virtual spaces out of large basis sets for transcorrelated (TC) calculations.
Kai: This new scheme, termed SVD-TC, performs post Hartree-Fock TC calculations in a compressed virtual orbital subspace obtained by projecting canonical virtual orbitals from a large basis set onto a smaller one through singular value decomposition (SVD).
Mira: This efficiency allows high accuracy from large basis sets while avoiding the composite nature of reference correction methods, as the bottleneck steps incur cost only on the small virtual space calculation.
Kai: The SVD-xTC scheme replaces composite-energy formulations with a simple one-step TC workflow that retains an accurate reference orbital set throughout the TC workflow while compressing the virtual space to the size of a smaller basis.
Paper summary: Kai: So we're looking at this paper called Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set. Basically, they introduce a new way to make those big virtual spaces smaller for transcorrelated calculations, calling it SVD-TC.
Mira: Right, so the core idea here is that they take a large set of basis functions and use singular value decomposition to project them onto a smaller one. This lets them do post Hartree-Fock transcorrelated calculations in this compressed space instead of using the whole large set.
Lev: And what this means for real hardware is that if you have to run these heavy correlation steps, they only cost you as much as calculating on that smaller virtual space, which should make it way more efficient.
Kai: It builds on a finding they had earlier: the error in the reference basis converges slower than the correlation energy error. The SVD-TC scheme is designed to exploit that by compressing the space.
Mira: Exactly, and this compression allows them to avoid what they call the composite nature of reference correction methods. So instead of doing everything at once, they only do the heavy lifting on this small virtual space calculation itself.
Lev: From an engineering standpoint, that sounds like a big win because those bottleneck steps, like the integral calculation and the post Hartree-Fock correlation method such as CCSD(T), only have to deal with a smaller set of orbitals.
Kai: They take this new scheme and test it on something substantial: the G2-one set, which has fifty-five molecules with first- and second-row atoms <ref:2608.10658#pg2,55 molecules with first- and second-row atoms>. They use SVD-xTC-CCSD(T) to compute atomization energies for these molecules.
Mira: And the results they show are pretty strong. They compare their SVD-xTC results against near exact semistochastic heat bath configuration interaction reference values and experimental data.
Lev: Specifically, they found that this SVD-xTC method delivers chemical accuracy already with triple zeta basis sets when looking at atomization energies against the AVQZ basis set.
Kai: That's a big claim because achieving chemical accuracy usually requires much larger basis sets. They found the lowest mean absolute error of zero point four seven kilocalorie per mole in atomization energies against the SHCI plus PBE plus CV reference for that AVQZ basis set.
Mira: So, it seems this method is working well for these molecules, suggesting that the compression works effectively in practice to get good accuracy quickly.
Lev: They also showed how efficient these workflows are by providing timings for computing atomization energies on the G2-one set <ref:2608.10658#pg1>. On AVTZ they took forty-seven node-hours, and on AVQZ they took one hundred thirteen node-hours.
Kai: And Lev, what about convergence? The error metrics for the SVD-xTC minus PP CCSD(T) showed that the mean absolute error and the root mean square error in xTC CCSD(T) atomization energies converged on AVDZ and didn't get better when they went to AVQZ.
Mira: That suggests that basis set convergence is much faster with these ECPs, converging in AVTZ instead of needing a larger set for the final result.
Lev: That’s interesting because it points to how quickly the method settles, which is important when you're trying to make this run on actual hardware where resources are tight.
Kai: So, to sum up this paper, "Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set," they’ve created SVD-TC. This scheme compresses virtual spaces using SVD for post Hartree-Fock calculations in a smaller subspace obtained by projecting large basis sets onto smaller ones.
Mira: And this compression lets them perform these heavy steps on a much smaller calculation, avoiding the composite nature of reference correction methods entirely. It’s about making the bottleneck steps only cost as much as that small virtual space calculation.
Lev: So what this means for the world is that we can get benchmark-quality transcorrelated results on more complex molecules than before, using fewer resources because of this efficiency gain.
Kai: The authors also used the quadruple zeta results to look at the accuracy of pseudopotentials within the TC method and showed that this workflow provides faster basis set convergence than all-electron TC.
Mira: So they're pointing toward a future where we can use these optimized workflows to push chemical accuracy even further, potentially combining them with favorable Jastrow optimization for all-electron transcorrelated workflows.
Lev: The limitation the authors mentioned is that this scheme focuses on compressing the virtual space, and we have to see how robust this compression is when you move beyond the G2 set they tested.
Kai: Yeah, so it’s a promising way to handle the computational cost in transcorrelated methods without having to use massive basis sets from the start. This paper gives us a concrete example of how SVD-TC achieves that efficiency.
Conclusion: Kai: So we've been looking at how they used singular value decomposition to shrink virtual spaces for transcorrelated calculations on the G2 set, and now we need to wrap up what this means for us.
Mira: Basically, the title tells you they took these huge basis sets and compressed them down using singular value decomposition so they could do post Hartree-Fock transcorrelated methods without needing the full space.
Lev: From an engineering standpoint, that implies that the bottleneck steps in the calculation only have to work on this smaller virtual space, which should save a lot of computation time on real hardware.
Kai: Exactly. It's about taking a large set of molecular orbitals and projecting them onto a much smaller one so you can do those heavy correlation calculations faster.
Mira: And they showed that this approach is very effective, achieving chemical accuracy with triple zeta basis sets in atomization energies for the G2-one set molecules.
Lev: That's a big number to hear, so what do those timings actually look like when we run these workflows on real systems?
Kai: Well, they showed that for the G2-one set, using SVD-xTC and CCSD(T) with the AVQZ basis set took about one hundred thirteen node hours.
Mira: And they found that this method converges much faster than other reference correction methods when you look at how the error changes as you increase your basis set size.
Lev: That's interesting because it suggests that for real hardware, we might not need to go to even bigger basis sets right away to get good results with these transcorrelated methods.
Kai: So, the main idea is that SVD-TC offers a way to handle the computational cost in these calculations by being smarter about how much space it actually needs to work in.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T