Cheap quantum-chemistry methods get some energies systematically wrong. Better references exist, such as coupled-cluster or quantum Monte Carlo calculations, but they are too expensive to run routinely. A natural business model follows: a provider trains a correction model on expensive reference data and sells corrections as a service. The trouble is that the customers who need this most, such as drug and materials companies, will not send proprietary molecular structures to a third party, and the provider will not hand over its model. Homomorphic encryption (HE), which lets a server compute on encrypted numbers without seeing them, could resolve this standoff. It is usually dismissed for scientific workloads because it runs 4 to 7 orders of magnitude slower than ordinary code and handles only additions and multiplications.
This note argues that the main error being corrected has an unusually encryption-friendly shape. Standard density functional theory (DFT) makes an electron partly repel itself. This "self-interaction error" makes the energy curve too low wherever charge is split between atoms. To leading order, that error is a quadratic function of how fractional the charge distribution is. If the customer computes physically meaningful descriptors of this error locally in plaintext, the provider's correction becomes a low-degree polynomial. That polynomial can be evaluated under approximate HE in two or three multiplication levels, with no bootstrapping (the expensive operation that refreshes accumulated encryption noise). The idea is to put the physics in the client's plaintext preprocessing and leave only a shallow polynomial on the server.
Core Claim
The model is a delocalization-error correction stacked on dispersion-corrected DFT and evaluated under encryption. It has three parts.
The first part is the client's physics. The client runs ordinary semi-local DFT with a Grimme D3 dispersion term. D3 is a public, cheap, pairwise $-C_6/R^6$ correction with damping, and it supplies the long-range van der Waals attraction that semi-local functionals miss. The client also computes per-atom descriptors of delocalization error. The central descriptor is a finite-difference curvature of the energy with respect to fractional electron number, $\kappa = E(N+\delta)+E(N-\delta)-2E(N)$, alongside per-atom fractional occupations and D3 pair contributions.
The second part is the server's model. The server holds a correction trained on residuals between the client-side method and high-level references. That correction is restricted to a quadratic form in the encrypted descriptors.
The third part is the encryption itself. Evaluation uses the CKKS scheme, an approximate HE scheme for real numbers, at fixed depth.
The claim is twofold. First, a quadratic correction recovers most of the accuracy that an unrestricted plaintext correction model achieves on self-interaction-sensitive benchmarks. Second, CKKS evaluation, including the decryption noise needed for provable security, stays well below chemical accuracy (about 1 kcal/mol, or $1.6\times10^{-3}$ Hartree).
The quadratic restriction has a physical motivation, not just a cryptographic one. Exact theory requires the energy to vary linearly between integer electron numbers, whereas semi-local functionals produce a convex curve (On extending Kohn-Sham density functionals to systems with fractional number of electrons). On any interval, the leading deviation of a smooth convex curve from its chord is a quadratic in the fractional occupation $f$, of size proportional to $f(1-f)$ times the curvature, as a Taylor expansion about the integers shows. A correction of the form
$$\Delta E = \sum_{a} \left( w_0 + \mathbf{w}^\top \mathbf{d}_a + \mathbf{d}_a^\top Q\, \mathbf{d}_a \right)$$
therefore captures the leading delocalization error. Here $\mathbf{d}_a$ is atom $a$'s descriptor vector (curvature, occupations, local D3 terms), and $w_0$, $\mathbf{w}$ and $Q$ are the provider's learned parameters. The ciphertext-by-ciphertext multiplication depth is one; plaintext weight multiplications and rescaling bring the total to roughly two or three levels, and the sum over atoms is done with slot rotations.
Relation to Prior Work
The research file's candidate gaps frame the problem differently, and that framing is what this proposal rejects. They propose evaluating the entire interatomic potential or dispersion term under encryption, including exponential radial functions, $1/R$ and $1/R^6$ inversions, and damping. That route inherits the deep polynomial approximations and division-free reformulations that make encrypted code so slow (ML Confidential: Machine Learning on Encrypted Data). I could find no work that runs any part of DFT, dispersion corrections or learned interatomic potentials under HE. The difference here is architectural: every operation that depends only on the client's secret and on public physics stays in plaintext on the client. Encryption protects only the step where client data meets the provider's secret.
Several lines of work supply the method. Encrypted-inference systems show the cost structure and the polynomial toolkit. CryptoNets replaced activations with $x^2$ and paid a 250-second single-query latency. SAFENet assigns polynomial degree channel by channel. PolyNeXt builds activation-free networks from elementwise products, but at large sequential depth. All three approximate a network originally designed with non-polynomial operations. Here, by contrast, the target function is argued to be near-quadratic from physics, so no approximation of a deeper model is needed.
The learned-functional work defines the accuracy ceiling. Kohn-Sham Regularization trains a neural exchange-correlation functional inside a differentiable self-consistent loop and removes self-interaction error in one-dimensional H$_2^+$. It motivates learning self-interaction fixes. However, it runs in plaintext and in one dimension, and it places the network inside an iterative loop that HE cannot run, because ciphertexts allow neither data-dependent branching nor stopping rules. This proposal deliberately keeps the self-consistent loop on the client. The cost is that the correction is applied after the calculation rather than self-consistently.
The dispersion literature constrains the design. Statistical Analysis of Semiclassical Dispersion Corrections shows that D3 parameter choices can swing the dispersion energy of the C$_{60}$ buckycatcher by 63.47 kcal/mol. Calculation of Ligand Dissociation Energies in Large Transition-Metal Complexes finds that PBE0-D3 matches local coupled cluster on WCCR10 to a mean absolute deviation of 4.3 kJ/mol, while both sit about 30 kJ/mol from experiment. Two consequences follow. The correction must be trained on residuals with respect to a fixed, declared D3 variant. And on WCCR10 it should target coupled cluster rather than the disputed experiments.
Securing Approximate Homomorphic Encryption Using Differential Privacy sets the precision constraint. It shows that CKKS decryption leaks the key unless noise of about $\kappa/2$ bits is added for $\kappa$-bit security. It also shows that bounds computed at runtime break security, so the error bounds must be fixed in advance.
Research Program
The work has four components.
The first is the descriptor layer. Building on the fractional-occupation formalism, which justifies Aufbau filling for fractional electron numbers, the client runs PBE-D3(BJ) on the system and on its $N\pm\delta$ variants. It extracts per-atom occupations, local curvature estimates and per-atom D3 sums. The cost is a few extra self-consistent calculations on the client.
The second is the correction model. Quadratic models of increasing rank in $Q$ are trained on residuals against references. They are compared with unrestricted plaintext neural correctors on the same descriptors. This comparison measures the "polynomial tax" in accuracy.
The third is the encrypted pipeline. The descriptors are packed one atom per slot, and the correction is evaluated in CKKS with parameters chosen statically. Differential-privacy decryption noise is added following Li et al. The procedure records the bits of usable precision left after that noise and compares them with the roughly $10^{-3}$ Hartree target.
The fourth is an honest threat model. HE hides the client's descriptors from the server. It does not stop a client from reconstructing the provider's model through repeated queries, and with a quadratic model this could require relatively few queries. This model-extraction risk is measured, not ignored.
The data come from the benchmarks in the file. Training uses the small-molecule benchmark collection GMTKN30, on which D3 was originally calibrated, plus S66. Testing uses data held out from training. S22 covers neutral non-covalent complexes. The charged-dimer, salt-bridge and water-cluster sets from the DFTB3/CPE-D3 study probe cases where delocalization error is large. WCCR10 covers transition-metal complexes against DLPNO-CCSD(T). The silicon self-interstitials in a 16+1-atom cell, against diffusion Monte Carlo, test extended systems where semi-local DFT underestimates formation energies by delocalizing charge.
Baselines
The naive reference point is the client's own PBE-D3 calculation with no correction. Every other method is reported as an improvement on it. On the silicon interstitials, the target values are the diffusion Monte Carlo formation energies of 4.4(1) eV for the split-110 site, 5.1(1) eV for the tetrahedral site and 4.7(1) eV for the hexagonal site (Accuracy of quantum Monte Carlo methods for point defects in solids). The file reports only that LDA and GGA fall well below these values; it gives no exact PBE numbers.
The strongest public competitor is a hybrid functional with dispersion, PBE0-D3, run by the client with no service at all. It is the strongest competitor because it is what a privacy-conscious customer would actually do. On WCCR10 it reaches a mean absolute deviation of 4.3 kJ/mol against coupled cluster (Husch, Freitag, & Reiher). For charged dimers, the best reported low-cost number is DFTB3/CPE-D3 at an RMSD of 1.49 kcal/mol (Improving intermolecular interactions in DFTB3). DFT+U, which adds on-site Hubbard terms (DFT+U within the framework of linear combination of numerical atomic orbitals), is the natural public competitor for the silicon cells.
Fairness requires the same geometries, basis sets and D3 variant for every method. Baselines receive no tuning against the test sets, and the corrector receives none either. Beating these baselines means that encrypted PBE-D3 plus the correction has lower test error than client-side PBE0-D3 on WCCR10 and the charged sets, at lower client compute. On WCCR10 this means a mean absolute deviation below 4.3 kJ/mol.
Two ablations isolate the key ingredient. The first removes the curvature and occupation descriptors and keeps only geometric and D3 descriptors. It tests whether delocalization-aware inputs are what make a quadratic model enough. The second compares the plaintext quadratic, the plaintext unrestricted network and the encrypted quadratic. Together they separate the cost of restricting the model's form from the cost of encryption noise.
Why not just…?
Why not just encrypt the coordinates and evaluate the whole potential under HE? This is the obvious design, and it buys nothing. Interatomic distances, D3 sums and radial functions use only public functions of the client's own data, so they can be computed in plaintext before encryption without revealing anything. Encrypting them adds inversions and transcendental approximations, which are exactly what produce the 6 to 7 orders of magnitude slowdown reported by ML Confidential. The proposal reports the multiplication-level count for both designs to make the comparison explicit.
Why not just use interactive secure computation, or a trusted hardware enclave? Interactive protocols such as SecureNets support exact nonlinear functions, but they require the client to stay online across communication rounds, which is a poor fit for asynchronous batch screening. Enclaves give up the cryptographic guarantee. For a depth-two polynomial, though, the gap in speed is likely small. The latency and bytes transferred per query for non-interactive CKKS are therefore measured alongside the interactive alternative. If the interactive approach turns out cheaper on realistic networks, it will be reported as such.
Why not just run a hybrid functional or DFT+U locally and skip the service? This is the real threat. If client-side PBE0-D3 is as accurate as the encrypted corrector, the service has no purpose. The head-to-head comparison against PBE0-D3 and DFT+U in the Baselines section settles it. The proposal succeeds only if it wins on accuracy, on cost per structure, or on both.
Proposed Experiments
The first decisive check is the polynomial tax. It asks whether the plaintext quadratic model comes within about 0.3 kcal/mol of the unrestricted plaintext network on S22 and the charged sets. If it does not, the physical argument for a quadratic form fails beyond leading order. The fallback is to add cubic terms on the highest-curvature atoms only, assigning degree selectively in the spirit of SAFENet.
The second check is precision. It measures whether CKKS with static parameters and differential-privacy decryption noise sized for 128-bit security keeps the added error below $10^{-4}$ Hartree on every test system. A failure here is a direct, publishable measurement of the precision-versus-security tension raised by Li et al.
The third check is generalization. On the silicon interstitials, the question is whether a correction trained only on molecules moves PBE-D3 formation energies toward the diffusion Monte Carlo values. On WCCR10, the question is whether the correction beats PBE0-D3's 4.3 kJ/mol against coupled cluster.
The final measurement is the attack cost. It records how many queries a client needs to reconstruct $Q$ to within its test-set error, which makes clear what this service protects and what it does not.
References
- On extending Kohn-Sham density functionals to systems with fractional number of electrons
- ML Confidential: Machine Learning on Encrypted Data
- CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy
- SAFENet: A Secure, Accurate and Fast Neural Network Inference
- Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models
- Kohn-Sham Equations as Regularizer: Building Prior Knowledge into Machine-Learned Physics
- Statistical Analysis of Semiclassical Dispersion Corrections
- Calculation of Ligand Dissociation Energies in Large Transition-Metal Complexes
- Securing Approximate Homomorphic Encryption Using Differential Privacy
- Accuracy of quantum Monte Carlo methods for point defects in solids
- Improving intermolecular interactions in DFTB3 using extended polarization from chemical-potential equalization
- DFT+U within the framework of linear combination of numerical atomic orbitals
- SecureNets: Secure Inference of Deep Neural Networks on an Untrusted Cloud