Searching Four-Phonon Surrogates on a Three-Phonon Budget

Neural networks can now stand in for the most expensive step in predicting how well a crystal conducts heat. That step is computing how often lattice vibrations (phonons) scatter off one another when four phonons, rather than the usual three, interact at once. One recent surrogate cut a combined three- and four-phonon calculation from about 1,285 to about 20 hours while keeping thermal conductivity within 5% of the exact answer. Its network shape, however, was chosen by hand. The obvious next move is to run neural architecture search, the automated design of network shape, on this surrogate. That move runs into a problem that is specific to this setting. In vision, the cost of architecture search is GPU time spent training candidates. In this setting, the dominant cost is the training labels: one four-phonon calculation for silicon takes over 7,000 CPU hours. An architecture search that needs many four-phonon labels to tell good candidates from bad ones can cost more than the surrogate saves.

This note proposes ranking candidate surrogate architectures on cheap three-phonon scattering data, scored by an error measure weighted by each mode's contribution to thermal conductivity, and only then spending scarce four-phonon labels on the few finalists. The claim to test is that this low-cost ranking agrees with the true ranking (accuracy of the predicted thermal conductivity once each architecture is trained on four-phonon data) better than zero-cost proxies or few-epoch training, and better than a small four-phonon validation set of equal CPU cost. The work charges every CPU hour of the search against the savings it produces. It also treats "a tuned hand-designed network is good enough" as a live outcome, not a straw man.

Background and the gap

The physical motivation is that four-phonon scattering is not a small correction. In penta-PdP₂ and penta-PtP₂ monolayers, adding it cut room-temperature lattice thermal conductivity by about 80% and roughly quadrupled the predicted thermoelectric figure of merit (Impact of Four-Phonon Scattering on Thermal Transport and Thermoelectric Performance of Penta-XP2 (X = Pd, Pt) Monolayers). This work establishes that the expensive physics matters. It does not offer a cheaper way to compute it.

The method we build on is Fast and accurate machine learning prediction of phonon scattering rates and lattice thermal conductivity. It trains fully connected networks to predict per-mode scattering rates from four inputs: phonon frequency, wave vector, polarization eigenvectors, and group velocity. It log-transforms the rates, which span orders of magnitude, and transfers what it learned on three-phonon data to four-phonon data through a "virtual phonon" construction. Its architecture was fixed by ad-hoc trial. Its eigenvector input grows as three times the number of atoms in the cell, so one network cannot serve crystals of different sizes. Our proposal keeps its data pipeline and loss, and searches over the network around it.

Prior architecture work on transport properties stops at hyperparameter tuning. Predicting thermoelectric properties from crystal graphs and material descriptors used Hyperband to tune fixed graph networks, descriptor networks, and XGBoost on a power-factor dataset whose labels assume one constant relaxation time for all electrons. The descriptor network won (20.70% relative mean absolute error, against 55.50% for the graph network). That result also warns us that sophisticated architectures can lose badly on transport targets.

The architecture-search literature supplies the cost-saving tools, and it also limits what to expect from them. Zero-cost proxies score untrained networks from a single minibatch. Evolved proxies reach Kendall τ = 0.65 on NAS-Bench-201 CIFAR-10 (EZNAS). Across 28 tasks, however, these proxies are volatile and biased toward parameter count (NAS-Bench-Suite-Zero). Budgeted estimation, meaning ranking architectures after cheap partial training, gave a 1,000× search speedup in vision (Rethinking Performance Estimation in Neural Architecture Search). Our "budget" is a cheaper physics fidelity rather than fewer training epochs, which is the main difference from that work. Most importantly, NAS-Bench-360 found that architecture rankings on CIFAR-100 have almost zero correlation with rankings on scientific tasks, and that tuned baselines often match search under constrained budgets. We therefore do not assume search helps. We measure whether it does, net of cost.

Core claim

We make two claims, each stated so it can fail.

The first concerns ranking. Two choices together should rank surrogate architectures for four-phonon data well. The first is the training data: three-phonon scattering rates, which cost far less per material than four-phonon rates. The second is the scoring rule: a conductivity-weighted error, which is specific to this problem.

Thermal conductivity in the relaxation-time picture sums, over phonon modes $\lambda$, the quantity $c_\lambda v_\lambda^2 \tau_\lambda$. Here $c_\lambda$ is the mode heat capacity, $v_\lambda$ the group velocity, and $\tau_\lambda = 1/\Gamma_\lambda$ the lifetime, the inverse of the scattering rate $\Gamma_\lambda$. A uniform error on log-rates over-weights modes that carry almost no heat. We therefore score candidates with

$$E_\kappa = \frac{\sum_\lambda c_\lambda v_\lambda^2 \,|\hat\tau_\lambda - \tau_\lambda|}{\sum_\lambda c_\lambda v_\lambda^2 \tau_\lambda},$$

where $\hat\tau_\lambda$ is the predicted lifetime. This weighted error bounds the relative conductivity error by the triangle inequality, so it measures the quantity the surrogate exists to get right.

The second claim concerns cost. Let $C_{\text{label}}$ be the CPU hours spent on first-principles labels used for search, $C_{\text{search}}$ the CPU and GPU hours spent training candidates, and $C_{\text{saved}}$ the hours saved by deploying the chosen surrogate over a target set of $M$ materials instead of the manual surrogate. Search is worthwhile only if

$$C_{\text{label}} + C_{\text{search}} < C_{\text{saved}}(M).$$

We report this break-even point $M$ explicitly for every search strategy, which no prior surrogate or architecture-search paper in the file does.

Research program

We would build a small tabular benchmark: a precomputed table of the results of every architecture in a fixed search space, in the spirit of NAS-Bench-360's precomputed Darcy Flow table. The search space is deliberately modest so that every candidate can be trained exhaustively. It covers network depth and width, activation function, residual connections, and the input encoding of the eigenvectors. The encoding is either the raw 3N-dimensional vector, as in the npj surrogate, or a per-atom shared encoder pooled across atoms. Pooling makes the network independent of cell size, which addresses that paper's transferability limit. It is the one structural choice we expect to matter most, and the search makes it a measured question rather than an assumption. The space contains on the order of a thousand candidates.

Labels come from first-principles three- and four-phonon scattering-rate calculations on Si, MgO, and LiCoO₂, the materials used in the npj surrogate study. Penta-PdP₂ and penta-PtP₂ serve as strongly anharmonic two-dimensional test cases. Four-phonon rates are computed on a sampled subset of modes, as the npj workflow does, so the label budget is finite and recorded.

Every architecture is trained on four-phonon data, with three-phonon pretraining and virtual-phonon transfer as in the npj work. Each trained model is then used inside the relaxation-time solver to produce thermal conductivity. The resulting conductivity error against full first-principles values is the ground-truth ranking. Every cheaper estimator is then compared against this ranking at matched CPU-hour cost.

Baselines

The naive reference point is no search: the hand-designed fully connected network of Fast and accurate machine learning prediction of phonon scattering rates and lattice thermal conductivity, retrained on our labels with its published loss and transfer scheme. That paper reports thermal conductivity within 5% of first principles on Si, MgO, and LiCoO₂, and a 70× workflow speedup. This is the number any searched architecture must beat. Doing so means a lower conductivity error on held-out materials, especially the penta monolayers, at equal or lower inference cost, with the search cost included in the break-even calculation. Beyond the reproduced npj model, we also report the full first-principles four-phonon calculation itself. It gives the cost ceiling and the accuracy reference.

The strongest competitor is not another search method but Hyperband tuning of the npj network's width, depth, and learning rate, as Predicting thermoelectric properties from crystal graphs and material descriptors did for its models. NAS-Bench-360 found tuned baselines of this kind often match search when budgets are limited. Hyperband gets the same label set, the same CPU and GPU hours as our full pipeline, and the same evaluation on held-out materials. Search beats it only if the selected architecture's conductivity error is lower by more than the seed-to-seed spread, measured over five training seeds.

For the ranking question, the competitors are four estimators. The first is zero-cost proxies (synflow, nwot, grad_norm, and their XGBoost combination, following NAS-Bench-Suite-Zero), computed on four-phonon minibatches. The second is few-epoch budgeted training on four-phonon data, following the budgeted-estimation work. The third is a small random four-phonon validation subset whose CPU cost equals that of our three-phonon ranking. The fourth is our three-phonon, conductivity-weighted ranking. For reference only, since the domain differs, EZNAS reports Kendall τ = 0.65 on NAS-Bench-201 CIFAR-10. No reported result was found on Lacuna for proxy or budgeted rank correlation on phonon scattering surrogates. An estimator wins if it has higher Kendall τ against the ground-truth conductivity ranking, together with higher overlap with the true top 10%, the bracket where proxies are known to fail.

Two ablations isolate the key ingredients. The first ranks on three-phonon data with uniform log-rate error instead of the conductivity weighting. The second ranks on four-phonon subsets with the weighting. Together they separate the effect of the cheaper fidelity from the effect of the physics-aware score. A tabular reference uses XGBoost on the same per-mode descriptors, trained with the same budget.

Why not just…?

Why not just Hyperband-tune the existing network? This is the real threat, and NAS-Bench-360's "buy vs. build" result predicts it may win. Our experiment runs it at equal total cost. If it matches the searched architecture, we report that architecture search does not pay for phonon surrogates. That is a useful negative result, and the three-phonon ranking study stands on its own because it would equally cheapen Hyperband's own evaluations.

Why not just use zero-cost proxies, which need no training at all? They cost nothing in labels, which is attractive here. However, NAS-Bench-Suite-Zero shows they swing between tasks and track parameter count, and their operators were designed for convolution-batchnorm-activation blocks rather than per-mode regression. They are a baseline in our ranking comparison. If they rank the top decile as well as the three-phonon approach, the simpler method wins and we say so.

Why not just predict thermal conductivity end-to-end from crystal structure, or use XGBoost? The npj study reports that end-to-end models often show 30–100% errors, against under 5% for per-mode surrogates. The thermoelectric study found graph networks failing on transport targets. XGBoost on per-mode descriptors is still included, because the JARVIS-Leaderboard finds classical descriptor models within 20% on nearly a third of AI tasks (JARVIS-Leaderboard).

Why not just hold out a few four-phonon labels for validation? This is the most direct competitor to three-phonon ranking. It is run at matched CPU hours. Three-phonon data wins only if it buys enough extra modes to give a more reliable ranking than a tiny four-phonon set. Whether it does depends on whether which architectures work well depends on the scattering order, which is exactly what we measure.

Decisive checks

The proposal succeeds on two conditions. First, the three-phonon conductivity-weighted ranking reaches higher Kendall τ and top-10% overlap than every matched-cost alternative across all five materials. Second, the selected architecture beats both the npj network and Hyperband on held-out conductivity error, with a finite break-even $M$. The most likely interesting outcome is that the pooled per-atom encoding, rather than depth or width, drives the gains, which would point future surrogate design toward structure, not search.

The proposal fails informatively if architecture rankings change between three- and four-phonon data. That would mean four-phonon physics needs different networks, a finding of its own. It also fails informatively if no search beats tuned Hyperband net of cost.