acceptodds
Under review as a conference paper at ICLR 2027

Scaling Density Functional Theory with Gaussian Splatting

Abstract

Density functional theory (DFT) strikes a practical balance between accuracy and computational cost in many problems of computational chemistry and materials science. However, many DFT calculations are limited by fixed atom-centered basis sets, which dictate how accuracy and cost scale with system size. We propose **Gaussian Splatting for Density Functional Theory (GS-DFT)**, which represents molecular orbitals as a cloud of Gaussians whose positions, shapes, and mixing coefficients are optimized jointly by gradient descent to minimize the energy without training data. Conceptually, GS-DFT is 3D Gaussian splatting with the renderer replaced by quantum mechanics. We introduce two key solver components: adaptive density fitting with screening for efficient evaluation of two-electron integrals, and a regularized differentiable orthogonalization of the molecular orbitals. Empirically, the optimized basis reaches the accuracy of the largest conventional basis sets with a fraction of the parameters, converging systematically in energy, density, and nuclear forces. At equal parameter count, it captures the stretched-bond and anion physics that fixed bases only recover with specialized basis augmentation. The resulting solver exhibits quadratic peak memory scaling in the cloud size, allowing us to simulate systems of up to atoms ( electrons) without any modifications at triple-zeta scale using a single four-GPU node.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
gaussian splatting, density functional theory, ai for science, quantum chemistry, scaling laws

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_6RRaOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a novel method of representing molecular orbitals as learnable floating Gaussians and variationally optimizing this representation with a differentiable KS-DFT implementation. Alongside the new orbitals, the paper also introduces two solver steps that save computational effort and allow for stable gradient propagation, respectively. GS-DFT is then tested on a range of molecular experimental setups, including significantly large structures and across several functionals, demonstrating strong accuracy and favorable memory scaling. Notably, the new adaptive basis functions do not have to be expert-tuned and generalize well to difficult systems such as diffuse ones.

Critical strengths

Clear and detailed presentation:

The paper is well written and presented. Every method is motivated and the paper contains extensive results and discussions. All aspects required to frame and understand the problem are clearly laid out by the authors, leaving little to no confusion after reading it.

Promising method:

The presented method is very promising. Instead of orbital-free DFT, this method still retains the typical variational KS-DFT formulation while the basis set itself is system-agnostic. Since choosing the right basis set is often a non-trivial task for difficult systems, this method could present an all-inclusive approach. The combination of an adaptive basis and efficient density fitting represents a particularly interesting direction for improving the scalability of electronic-structure calculations.

Critical weaknesses

Computational cost:

As noted by the authors in Appendix D.4, the method is significantly slower in wall-clock time than conventional SCF calculations because jointly optimization of the basis and orbital coefficients introduces computational overhead. I do not believe this undermines the methodological contribution, but it limits the immediate applicability of the method as a replacement for conventional DFT solvers.

On a related note, the largest-system experiments have favorable memory scaling on calculations involving thousands of atoms. However, these results primarily establish computational feasibility rather than accuracy at this scale. I recognize that performing fully converged conventional DFT calculations on the largest systems is likely prohibitively computationally expensive, but it would have been instructive to see an "accuracy versus system size" curve of the error per atom extending as far as computationally feasible into the large-system regime (e.g., 50, 100, 200, 300, 500 atoms or whatever is feasible). Even comparisons against conventional DFT at a fixed functional and basis set level would provide useful evidence.

However, neither limitation substantially changes my overall positive assessment of the paper.

Questions for authors
  1. Could GS-DFT be extended to open-shell systems, for example through unrestricted KS-DFT? Or are there reasons why this would be difficult?

  2. Is there a reason you did not chemically inform all initializations, e.g., your results in Figure 3? Based on the ablation in Appendix E.3, it seems like the best approach. Did you attempt to pre-fit the splats in a superposition-of-atomic-densities way?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6RRa

We thank the referee for their feedback. Our response can be found below.

  1. The answer is yes. We can extend the method and the implementation to open-shell systems by jointly optimising a coefficient matrix for each spin channel. This can be easily wired throughout the code. We strongly believe this is a necessary next step to deal with more complex, relevant systems. However, due to time constraints, we must defer this direction and the analysis of their optimisation dynamics to further studies.

  2. We wanted to show that even random initialisation yields correct results as we first focus on correctness, leaving the convergence cost as a future problem. Furthermore, we indeed thought about doing many different initializations but refrained from tackling them all to better focus on the other parts of the method. It is nonetheless a necessary future direction, as a good initialization dramatically reduces the number of steps necessary for convergence. This is a fantastic insight from the referee.

Regarding accuracy versus system size, Figure 9c in Appendix F.3 reports the error per atom at each system's cc-pVTZ size from 7 to 196 atoms: splats are 2 to 6 times more accurate than cc-pVTZ on every C/H/N/O system. Figure 6 shows the full accuracy scaling of the two largest, at 168 and 196 atoms. While a full accuracy scaling analysis of our largest systems is unfeasible, we provide such analysis for a bridging, intermediate scale.

While there were not any major concerns on my part to begin with, I think the authors have answered well for themselves on the questions that I and other reviewers had.

This only reinforces my positive assessment of the paper, which I believe is timely and presented well.

Reviewer_5ea5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes a new computational strategy for density functional theory for molecular systems, named Gaussian Splatting for Density Functional Theory. It replaces the conventional representation using a fixed atom-centered basis with a set of Gaussian basis functions with flexible parameters. With the DFT total energy as the objective function, this work adopts a set of strategies to find the optimal parameters for each system separately while maintaining the required physical properties. To reduce the overall computational complexity of Coulomb interactions, this work introduces an adaptive density-fitting strategy. The locality and analytical expressions of the Gaussian functions ensure the low computational cost of density fitting. All of the strategies above are supported by detailed mathematical proofs, as well as a clear step-by-step implementation.

Numerical experiments evaluate several chemical observables (i.e., energies, density, forces, and ionization potentials) using different types of functionals, suggesting potentially broad applicability. Calculations on a set of large systems are presented to illustrate the overall reduction in complexity.

Critical strengths
  1. A detailed description for the proposed strategy GSDFT, supported by theoretical mathematical analysis and multi-GPU implementation.

  2. The experiments are conducted to verify both accuracy and complexity, both reaching significantly behavior.

Critical weaknesses
  1. Algorithmic novelty is not clearly positioned relative to prior DF work.

    The abstract highlights adaptive density fitting (DF) and regularized differentiable orthogonalization as key components, but Algorithm 1’s outer-loop refresh of the screened pair list/auxiliary basis and inner-loop optimization with the auxiliary representation fixed make it unclear how memory benefits should be attributed specifically to DF; the paper also does not clearly compare the proposed adaptive DF strategy and update procedure with existing DF approaches or establish the corresponding complexity.

  2. The connection between small-system accuracy and large-system scalability is insufficiently established.

    The large-system demonstrations show computational potential but are not supported by a representative intermediate-scale benchmark with explicit convergence and accuracy comparisons against a suitably converged reference; the middle-size results mainly characterize cost and optimization progress rather than solution accuracy, and although Appendix A provides error analysis, the tightness of the bounds is not sufficiently characterized.

  3. The numerical comparisons are not fully validated or controlled.

    Table 3 compares GS-DFT with extrapolated memory estimates for density-fitted cc-pVTZ calculations without verifying that cc-pVTZ produces accurate results for the given systems or establishing the accuracy/runtime/memory trade-off under comparable conditions; Table 10 lists software capabilities rather than the configurations needed for a fair comparison with DF-type strategies; it is unclear whether density fitting was enabled for PySCF; and a benchmark on systems containing tens to a few hundred atoms comparing GS-DFT with an established density-fitted solver under the same functional and geometry is missing.

Questions for authors
  1. How does the proposed adaptive DF strategy and update procedure differ from existing DF approaches, and what is the corresponding complexity comparison?

  2. Given Algorithm 1’s structure, how should the memory boosting/reduction be attributed specifically to DF rather than to the orthogonalization or other solver components?

  3. Can the authors provide a representative intermediate-scale benchmark with explicit convergence and accuracy comparisons against a suitably converged reference, and better characterize the tightness of the error bounds in Appendix A?

  4. In Table 3, did the cc-pVTZ calculations produce accurate results for the given systems, and can the authors establish the corresponding accuracy/runtime/memory trade-off under comparable conditions?

  5. For Table 10, which configurations should be used to make a fair comparison with DF-type strategies, and was density fitting enabled for PySCF?

  6. Can the authors add a benchmark on systems containing tens to a few hundred atoms comparing GS-DFT with an established density-fitted solver under the same functional and geometry to evaluate practical advantages at a reasonable accuracy level?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5ea5

We thank the referee for their feedback. Our response can be found below.

  1. The core difference lies in the fact that the conventional density fitting (DF) is a low-rank approximation of a tensor in a fixed basis, while our adaptive density fitting is a low-rank approximation of a changing basis. We construct this approximation as a function of the splat set every KK steps as the individual splat movements change the density little. Regarding the complexity comparison, DF is O(B2⋅Baux)O(B^2 \cdot B_\text{aux}) with BB being the number of basis functions and BauxB_\text{aux} the number of auxiliary basis functions. Usually, BauxB_\text{aux} is 2 to 4 times BB; thus, DF is O(B3)O(B^3). Meanwhile, with MM the number of splats, Maux=2MM_\text{aux} = 2M auxiliary functions and kk retained neighbours per splat after screening, ADF costs O(MkMaux+Maux2)O(M k M_\text{aux} + M_\text{aux}^2) per step, plus the O(Maux3)O(M_\text{aux}^3) factorization of the auxiliary metric amortized across KK refresh steps (Table 7). As shown by the accuracy scaling analysis (Figure 9b in Appendix F.3), splats reach 1 mHa with 5 to 11 times fewer basis functions than the cc-pVXZ ladder, so M<BM < B at a given accuracy, and increasingly so as the system size grows.

  2. The dominant memory scaling of DFT comes from the two-electron integrals (Dunlap et al., 1979; Vahtras et al., 1993), and we further refer back to our previous answer for a complexity comparison. Furthermore, the other parts maintain the same complexity as their counterparts in previous DFT solvers.

  3. We refer the referee to our answer for question 1 of reviewer jqdY and to Figures 5 and 6, and Figures 8 and 9 in Appendix F.3. We show the accuracy scaling on eight new systems ranging from small (7 to 74 atoms) to medium sized (168 and 196 atoms). Each system is measured against its own complete basis set (CBS) limit, extrapolated from its cc-pVXZ ladder up to cc-pVQZ or cc-pV5Z, which serves as the converged reference. We would like to further point out that despite the previous characterisation of small and medium, systems of more than 100 atoms are considered large to very large in DFT. This is especially the case when doing calculations at TZ accuracy scale.

  4. We would like to clarify that the cc-pVTZ calculations in Table 3 are not feasible, so they were not run: the GTO memory column is an estimate extrapolated from measured runs. The reference energies are approximate values from the FMO database (Takaya et al., 2021), calculated using the fragment molecular orbital (FMO) method with Hartree-Fock and the 6-31G(d) basis. This approximate method is necessary at this scale as conventional DFT is unfeasible. As these experiments are not feasible, we cannot compute an accuracy, runtime, and memory comparison under identical conditions at this scale. Table 3 has been re-measured with a more efficient implementation to further emphasise this. For accuracy as a function of system size, Figure 9c in Appendix F.3 reports the error per atom at each system's cc-pVTZ size from 7 to 196 atoms, where splats are 2 to 6 times more accurate than cc-pVTZ on the C/H/N/O systems.

  5. When computing the baseline values for PySCF, density fitting was enabled. Furthermore, the baselines were run on identical hardware and with density fitting when available.

  6. We refer back to our answer of question 3 where we show the accuracy scaling of eight new systems with fixed geometries and XC functionals, amongst them two of intermediate size relative to our experimental validations. All Gaussian ladders of these systems were corroborated with GPU4PySCF (Wu et al., 2025).

References

  • Dunlap, B. I., Connolly, J. W. D. and Sabin, J. R. On some approximations in applications of Xα theory. J. Chem. Phys. 71, 3396–3402 (1979). https://doi.org/10.1063/1.438774
  • Takaya, D. et al. FMODB: The world's first database of quantum mechanical calculations for biomacromolecules based on the fragment molecular orbital method. J. Chem. Inf. Model. 61, 777–794 (2021). https://doi.org/10.1021/acs.jcim.0c01062
  • Vahtras, O., Almlöf, J. and Feyereisen, M. W. Integral approximations for LCAO-SCF calculations. Chem. Phys. Lett. 213, 514–518 (1993). https://doi.org/10.1016/0009-2614(93)89151-7
  • Wu, X. et al. Enhancing GPU-acceleration in the Python-based simulations of chemistry frameworks. WIREs Comput. Mol. Sci. 15, e70008 (2025). https://doi.org/10.1002/wcms.70008

We thank the referee again for their review. We have posted a detailed response to each of their comments, with new experiments and revisions to the manuscript, marked in red. As the discussion period is drawing to a close, we would greatly appreciate it the referee could look at our response and let us know whether it addresses your concerns. We would be glad to answer any remaining questions.

Reviewer_jqdYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed Gaussian Splatting for Density Functional Theory, a method that combines the idea of floating Gaussian basis and treating DFT as an optimization problem. This paper also introduces several computational optimizations, such as adaptive density fitting, to make the approach computationally tractable. The results demonstrate that GS-DFT required fewer basis functions for some systems.

Critical strengths

This is a very interesting approach. Connecting DFT with 3DGS is new to the field. This paper opens a new perspective in the DFT area. Treating DFT as an optimization problem has a long history, but the previous attempts, such as FEGO, fail to scale up to a meaningful size of molecules. The computational optimization introduced by this paper finally makes this a working idea.

Critical weaknesses

This paper makes a relatively broad claim about scalability and efficiency, but the experiments don't fully support the claims. Specifically:

  1. One of the main contributions, "showing 2.3× better scaling for all systems," is computed based on three examples; it's not for all systems. Is it true that 2.3x still holds in a more diverse testing set?

  2. I also noticed that "GS-DFT at a matched count of M = nao = 24 ends 66.6 mHa above the cc-pVDZ reference". So the "fewer basis functions" claim doesn't work in some cases, and there is a crossover point in Figure 5, after which GS-DFT needs fewer functions. And for different molecules, that crossover point seems to be different. Could the author comment on where the crossover point is, in general?

  3. Could the author comment on the optimization step scaling? This paper mainly uses a fixed iteration budget, but in practice, we run SCF until DFT is converged. So the question is, does the large system require significantly more steps to converge?

  4. Could the author comment on the memory bottleneck part, "Adding devices shortens each device’s share of the work but does not reduce its peak memory." It seems that the largest system it can handle is capped by a single GPU's memory. Could the author comment on sharing memory across the devices? So larger molecules can be calculated. Can it handle a larger system than conventional DFT? In Table 10, only the CPU version of PySCF is compared. Could we use GPU4PySCF and specify the density-fitting settings?


And some minor weaknesses:

  1. In Figure 5, this paper compares two differences: ΔGTO=EGTO−ECBS\Delta_{\mathrm{GTO}}=E_{\mathrm{GTO}}-E_{\mathrm{CBS}} and ΔGS=EGS−EISC\Delta_{\mathrm{GS}}=E_{\mathrm{GS}}-E_{\mathrm{ISC}}. But EISCE_{\mathrm{ISC}} is not the same as ECBSE_{\mathrm{CBS}}. Since one key conclusion, 2.3x, is calculated from Figure 5. Could the author comment on the effects on these two extrapolated limits?

  2. DFT partly relies on error cancellation of the incomplete basis. The downstream property (i.e., experimental quantities) is not necessarily best at the CBS limit. This paper demonstrates that GS-DFT may converge faster to the CBS limit. Could the author comment on whether GS-DFT can converge to a better physical property with fewer basis functions?

  3. The connection from 3DGS to this paper is not very clear to me. Treating DFT as an optimization problem and using floating Gaussian bases are established. I would consider the contribution in this paper to be making this combination computationally competitive. From my reading, only a small portion is directly borrowed from 3DGS's idea: the quaternion rotations part. The rest contribution, including density fitting, differentiable orthogonalization, GPU distribution are more close to the DFT area. Thus, I would consider this paper to be "inspired by 3DGS", rather than treating 3DGS as a core computational toolkit.

  4. Using quaternion rotations and log scale is a valid and convenient choice. But is there an advantage to using such a representation for the Gaussian precision matrix? Could the author compare its performance with other possible parametrization choices?

  5. This paper demonstrate multiple GPU is possible. Could the author also discuss the parallel efficiency?

Questions for authors

See weaknesses.

7 comments
AuthorsOpenReview ↗

First part of our response to Reviewer jqdY

We thank the referee for their feedback. Our response can be found below.

  1. To further substantiate our claims, we have expanded the scaling analysis, adding eight systems from 7 to 196 atoms. Figures 5 and 6, together with Figures 8 and 9 in Appendix F.3, show the accuracy scaling of these new systems, supporting our previous empirical findings of at least 2.3×2.3 \times better scaling than the baselines. We have chosen these systems to cover a wide variety of chemical settings.

  2. In Appendix D.2, we state that our method does not beat the baseline at small basis function count MM. Expanding the accuracy scaling analysis with more systems (please see our response to the first question) shows that as the system size grows, our method outperforms the baseline sooner. We further show the empirical accuracy crossing of the methods in Figure 9a in Appendix F.3. When we adjust for system size, the crossover sits at 13.3±2.513.3 \pm 2.5 basis functions per atom on the C/H/N/O systems, with no trend from 16 to 196 atoms.

  3. It is a well-known fact in DFT that both SCF iterations and direct energy minimisation require more iterations as the system size scales (Payne et al., 1992; Kresse and Furthmüller, 1996; Lin and Yang, 2013). More parameters are required to describe larger systems in most parametrizations of molecular orbitals. While SCF does take fewer iterations to converge than direct energy minimisation, the latter takes cheaper steps that do not require diagonalising the Hamiltonian (Payne et al., 1992; Van Voorhis and Head-Gordon, 2002).

  4. The specific line in question concerns the sharding of one-electron integrals, which are replicated across devices for the multi-GPU experiments. The sharded terms are the Hartree and XC terms, as they dominate the cost. To show this, we have provided a parallel efficiency analysis in time and memory found in Figure 10 (a: time, b: memory) in Appendix F.4. We observe near-perfect efficiency in time for the larger systems, and an efficiency in memory of about 95% on two GPUs and about 70% on four. Furthermore, our method can indeed handle larger systems than conventional DFT due to its better memory scaling. Finally, we have added a new row in Table 10 with estimated memory usage of GPU4PySCF (Wu et al., 2025), using density fitting. We note that GPU4PySCF has a known issue when using density fitting (GPU4PySCF issue #589), as it moves the three-index density-fitting tensor to host memory once that tensor no longer fits in GPU memory. Thus, it may silently require the same large host memory as traditional CPU-bound DFT calculations.

References

AuthorsOpenReview ↗

Second part of our response to Reviewer jqdY

Continuing the previous answer:

  1. We agree with the referee that the convergence sharpness relative increase by a factor of 2.3 can be affected by the mismatch between the CBS and ISC limits. Figure 3 presents results against both references and their small numerical difference. Since CBS and ISC have been extrapolated from independent GTO and splat calculations, they may retain systematic biases of the given representation family, the extrapolation function, or even any finite-basis error cancellation brought up by the referee. We do not currently have an independent third representation under the same XC functional that would establish which extrapolated reference is more reliable in some absolute sense. Therefore, in Figure 5, the convergence sharpness is estimated from how each sequence approaches its own estimated limit, without assuming either estimate to be superior. Additionally, we will highlight that we are reporting the approximate observed slope, not a method guarantee, as suggested by Jensen, 2002, with an independently chosen extrapolation function.

  2. We cannot make general statements about "finite-splat" cancellation because accurate experimental data is not widely available, especially at the scale GS-DFT performs best at. As indirect evidence for matching reference values, Table 2 compares the Hartree-Fock estimates of the ionization potential from GS-DFT with the CCSD(T) and experimental values of the GW100 benchmark, supplementing numerical evidence in Figure 3 and Figure 4. We will clarify the scope in the manuscript: GS-DFT is designed to represent solutions of a given XC approximation efficiently, while any improvement in experimental predictions remains uncontrolled and dependent on the functional itself and the observable.

  3. Our method is indeed inspired by 3D Gaussian Splatting (3DGS) in computer vision. Mathematically, we could observe equivalences between the photometric loss and the energy loss, and the rendering process and the occupied subspace construction of the density. We use 3DGS as a convenient analogy to showcase the similarities and explain ideas. Nonetheless, we do not claim to do exactly the same in both methods. We have reworded our statement to further highlight this (Appendix B.4).

  4. As opposed to alternatives like matrix exponentials or Cholesky decompositions, our Gaussian covariance parametrisation was chosen for two reasons. First, it naturally parametrises the varied scales of the basis set in log-space. Second, it avoids expensive diagonalisation or Cholesky decompositions on each 3 by 3 covariance block. As an example, the matrix exponential parametrisation, which requires diagonalising the covariances, resulted in equally accurate results but was 1.27×1.27\times more expensive in compute while using equal memory.

  5. We refer back to our answer to question 4. As shown in the parallel efficiency analysis. Sharding yields near-perfect speedups for the larger systems, and an efficiency in memory of about 95% on two GPUs and about 70% on four.

References

  • Jensen, F. Polarization consistent basis sets. II. Estimating the Kohn–Sham basis set limit. J. Chem. Phys. 116, 7372–7379 (2002). https://doi.org/10.1063/1.1465405
AuthorsOpenReview ↗

Additional results on observables (Appendix F.5)

We thank the referee again for their questions. To further support our answers to questions 1, 5 and 6, we have added Appendix F.5, which repeats the observables analysis of Figure 3 on six of the new systems: SF$_6$, cubane, the adenine–thymine base pair, porphine, retinal and cholesterol. Figure 11 reports the energy, the density and the nuclear forces of both the splat clouds and the cc-pVXZ ladder.

  • Question 1. As in Figure 3, the observables converge as the clouds grow on every system. This extends the evidence beyond the energy scaling of Figures 5, 6, 8 and 9 to the density and the forces, on a wider range of chemistry.
  • Question 5. Every system is measured against both the CBS and the ISC limit, which extends the comparison of the two references in Figure 3 to six more systems.
  • Question 6. For the larger clouds, and from cc-pVTZ on, the remaining density and force errors of both representations lie within the error of the PBE functional itself. Beyond this point, a smaller representation error brings the result closer to the exact PBE solution, not to the true electronic structure. This supports our answer that GS-DFT is designed to represent the solution of a given functional efficiently, while the accuracy of physical predictions at this scale is set by the functional rather than by the basis.

We hope these additional results address the referee's remaining concerns, and we would be glad to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the referees for the thorough reading of our manuscript and the opportunity to further clarify and improve our method. We hope that our responses will clarify key contributions and answer their questions with new evidence. We have uploaded a revised manuscript, with all changes highlighted in red. We look forward to improving the GS-DFT method through open discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.