When engineers design a lattice of coupled resonators to host topologically protected modes, they care about several numbers at once. How tightly the protected mode sits at the edge, how wide the energy gap of the bulk is, how much the mode's frequency drifts under fabrication error, and how much chip area the lattice takes. Each of these numbers is expensive to measure, because robustness alone needs hundreds of disorder-randomized simulations per candidate design. Multi-objective Bayesian optimization handles exactly this setting. It is a sample-efficient search that fits a probabilistic model, called a surrogate, to past evaluations and uses that model to choose the next design. The newest versions of these methods also model how the objectives move together. In topological systems, however, that co-movement is not fixed. Within one topological phase, a larger bulk gap means a shorter decay length for the edge mode, so the gap and edge localization rise together. At a phase boundary the gap closes and the edge mode disappears. In a different phase, the tradeoffs between objectives can reverse.
This note proposes a correlated surrogate whose correlations between objectives switch according to the topological phase of the design. The phase is identified by a topological invariant, an integer such as a winding number or Chern number that labels the phase. This invariant can be computed in milliseconds from an idealized model of the design. Because topological invariants do not change under small perturbations, that cheap label stays reliable even when the idealized model misses parasitic effects that dominate the continuous objectives. In two sentences, the idea is this: compute the cheap, robust invariant for every candidate, and let it select which correlation structure the expensive objectives follow. We expect this to waste fewer expensive evaluations than surrogates that assume one global correlation structure or none.
Core Claim
Existing correlated multi-objective methods assume one correlation structure everywhere in design space. Pareto Frontier Learning with Expensive Correlated Objectives (CEIPV) uses multi-task Gaussian processes with a single coregionalization matrix, which is the matrix that sets how strongly each pair of objectives co-varies. BOtied fits one vine copula, a model that builds a joint distribution from a tree of two-variable dependence models, to the whole objective space. PSIPS assumes a covariance shared across arms in a linear bandit.
We claim that in physical design problems governed by topological phases, this stationarity assumption fails. The failure is structured: correlations change abruptly where the integer invariant changes. A surrogate gated by the invariant should therefore be both better calibrated and more sample-efficient. Shah et al. report that ignoring correlations that really exist can make optimization worse than random. By the same reasoning, imposing one averaged correlation on a landscape whose correlations flip sign should also mislead the search, in the regions near phase boundaries where the interesting tradeoffs live.
Let $x$ be a design vector, $f(x) \in \mathbb{R}^M$ the vector of $M$ expensive objectives, and $c(x) \in \{1,\dots,K\}$ the phase label computed from the idealized model. The proposed prior covariance between objective $i$ at $x$ and objective $j$ at $x'$ is
$$ \mathrm{Cov}\big(f_i(x), f_j(x')\big) = \big(A_{c(x)} A_{c(x')}^{\top}\big)_{ij}\, R_{c(x)c(x')}\, k(x,x'). $$
In this formula, $A_k$ is a learned $M \times r$ loading matrix for phase $k$, so $A_k A_k^\top$ is that phase's correlation structure among objectives. $R$ is a learned $K \times K$ positive semidefinite matrix with unit diagonal that controls how much information passes between phases. $k$ is an ordinary kernel over designs. The whole expression is a product of two positive semidefinite kernels on (design, objective) pairs, so it is a valid Gaussian process. It reduces to CEIPV's multi-task model when $K=1$ and to independent models per phase when $R$ is the identity.
Physics supplies an optional prior on the sign of entries in $A_k$. For gapped phases, the decay length scales roughly as the band velocity divided by the gap, so gap and localization should co-vary positively inside a nontrivial phase.
Research Program
The work has three parts.
A testbed for higher-dimensional topological design. We start from the integrated circuit of Fully Integrated Topological Electronics. That paper built a one-dimensional Su-Schrieffer-Heeger (SSH) chain, the simplest lattice with a protected edge mode, from six capacitively coupled LC resonators in 130 nm CMOS. It reached an edge-mode inverse participation ratio of 0.871 against roughly 0.5 for bulk modes, and a 0.05% frequency variation coefficient across 500 Monte Carlo runs at ±10% component variation. The inverse participation ratio measures how concentrated a mode is on a few sites. The paper designed the circuit by hand and names 2D and 3D lattices as the next step, which is where we pick up.
We build a parameterized simulator for three lattices of increasing difficulty. The first is a 1D SSH chain, used for sanity checks against Liu et al.'s numbers. The second is a 2D SSH lattice with corner and edge modes. The third is a 3D stack of SSH layers with interlayer couplings, which supports both gapped phases and nodal-line semimetal phases. Nodal lines are gap closings along curves in momentum space, the same family as the high-symmetry-line semimetals catalogued in Machine Learning on Multiple Topological Materials Datasets.
The expensive oracle simulates a finite lattice at netlist level, with a parameterized parasitic model (next-nearest-neighbor coupling capacitance, substrate loss, finite inductor quality factor) and 200 disorder draws per design. It returns $M=5$ objectives: edge or corner inverse participation ratio, relative bulk gap, frequency variation coefficient, total capacitance as a proxy for area, and edge-mode quality factor. The cheap oracle computes the winding number, Chern number, or nodal-line charge from the ideal Bloch Hamiltonian.
A stress test with a delicate phase. We add a tight-binding-only model of the multi-band Hopf insulator from Lapierre and Trifunovic. Its invariant is only meaningful while the number of bands is held fixed, and the authors stress that it is unstable to perturbations that add bands. Parasitic couplings to extra modes therefore make the cheap label genuinely unreliable here. This testbed probes whether the method fails gracefully when the cheap model's invariant disagrees with the true finite system.
The surrogate and acquisition. We fit the phase-gated model above by marginal likelihood, using low rank $r \le 2$ per phase to control the cubic cost that Shah et al. flag for latent-factor models. Acquisition uses Monte Carlo expected hypervolume improvement with joint posterior samples. Hypervolume improvement is the gain in the volume of objective space dominated by the current best designs. The Monte Carlo version handles correlated posteriors and remains differentiable, which sidesteps the derivative-free inner optimizer BOtied requires. We also run BOtied's copula-based acquisition on top of our surrogate, so that the effect of the surrogate is not confounded with the choice of acquisition function.
Baselines
The naive reference points are Sobol random sampling and NSGA-II, a standard evolutionary algorithm that evolves a population toward the set of best tradeoffs, each run at the same budget of 150 expensive evaluations. NSGA-II is also run at ten times that budget as an upper reference. Shah et al. report that CEIPV significantly outperformed NSGA-II on synthetic landscapes at correlation $\rho=\pm0.75$, but give no numeric gap. We will therefore measure this ourselves.
The independence baseline is a set of separate Gaussian processes with Monte Carlo expected hypervolume improvement. This is the standard configuration criticized in the file's discussion of Belakaria et al. and Wang et al.. ParEGO, which collapses objectives into random weighted sums, serves as the scalarization baseline. The strongest reported number on this axis is Shah et al.'s result on the boston neural-network task: CEIPV reached in 10 iterations the optimization level ParEGO needed 100 iterations to reach.
The strongest existing correlated competitors are a CEIPV-style stationary multi-task Gaussian process and BOtied. For BOtied, Park et al. report best scores on the 9-objective Car Cab problem, but the file gives no numeric value. For our topological design task there is no reported result found on Lacuna for any multi-objective optimizer.
The key ablation is a phase-gated model with $R$ fixed to the identity, meaning separate correlated models per phase that share nothing across phases. A second ablation randomly permutes the phase labels while keeping their proportions, which tests whether the gain comes from the invariant itself or merely from extra parameters.
All methods share the same 20 initial Sobol designs and the same 150-evaluation budget. Every surrogate gets the same kernel family and the same hyperparameter fitting schedule, and every acquisition optimizer gets the same number of restarts. Each comparison uses 20 seeds.
Beating the strongest baseline means two things. The method must reach 90% of the reference hypervolume in at least 30% fewer expensive evaluations than the stationary multi-task model, with non-overlapping 95% intervals across seeds, on at least two of the three circuit lattices. It must also give lower held-out joint negative log-likelihood, which measures how well the model's predictive distribution fits unseen data.
Why not just…?
Why not just use the invariant as a feasibility constraint and run standard multi-objective optimization inside the nontrivial phase? This is the most serious threat. If every Pareto-optimal design lies in a single phase, filtering out trivial designs captures most of the benefit, and the phase-gated covariance adds only complexity. We do not argue this away. Constrained optimization with the same cheap invariant runs as a dedicated baseline. The 3D stacked lattice is chosen because its tradeoffs plausibly span a gapped phase, where localization is high but area is large, and a nodal-line phase, where the structure is compact but surface modes are weaker. That gives a case where the front crosses a boundary.
We will report how many phases the reference Pareto set occupies on each testbed. If constrained optimization matches our method whenever the front lies in one phase, we will say so. The claim then narrows to fronts that cross phases, and the synthetic control below tests that narrowed claim directly.
Why not just run NSGA-II on a fast simulator? The ideal tight-binding model is fast, but the objectives that matter are not: robustness requires a few hundred parasitic-aware finite-lattice simulations per design, mirroring the 500-run Monte Carlo of Liu et al. At a fixed budget of 150 evaluations, NSGA-II barely completes a few generations. The 10× NSGA-II reference shows how much budget evolution needs to catch up.
Why not just use one stationary correlated model such as CEIPV or BOtied? These methods capture real correlation, which Shah et al. show matters. They average it over the whole design space, though. If correlations really flip across phases, a single coregionalization matrix or copula will be poorly calibrated near boundaries. The labelled-versus-shuffled ablation and the held-out likelihood test whether this matters in practice.
Why not just train a classifier on materials databases to predict the phase? He et al. show that static classifiers fall from 85.7% accuracy on internal validation to 71.8% on out-of-distribution compounds, and they only predict labels rather than optimizing continuous device parameters. In our setting the phase can be computed exactly from the idealized model, so no learned classifier is needed.
Proposed Experiments
The first decisive check uses a synthetic control built from Shah et al.'s correlated test functions. We modify them so the correlation is $+0.75$ on one side of a known boundary and $-0.75$ on the other, and we provide the side as the cheap label. The phase-gated model should clearly beat the stationary multi-task model here. On an unmodified stationary-correlation version, it should show no gain. A gain in the stationary case would mean the improvement comes from extra flexibility rather than the mechanism we claim.
The second check runs all baselines on the three circuit lattices. We report hypervolume against evaluation count, evaluations needed to reach 90% of reference hypervolume, the fraction of evaluations spent in phases that contain no Pareto-optimal points, and wall-clock time per iteration. As a sanity check, the 1D lattice should recover edge localization near Liu et al.'s 0.871.
The third check inspects the learned loading matrices $A_k$. We test whether the gap–localization correlation is positive in gapped nontrivial phases and changes sign or vanishes across boundaries, as the decay-length argument predicts.
The fourth check corrupts the phase labels at rates of 0%, 10%, 20%, and 30%, and uses the Hopf testbed, where parasitic bands make labels genuinely unreliable. The learned between-phase matrix $R$ should move toward all ones as labels degrade, so the model falls back toward the stationary baseline rather than below it. Performing worse than the stationary model at moderate label noise would count against the proposal.
References
- Pareto Frontier Learning with Expensive Correlated Objectives
- BOtied: Multi-objective Bayesian optimization with tied multivariate ranks
- Pareto Set Identification With Posterior Sampling
- Fully integrated topological electronics
- Machine learning on multiple topological materials datasets
- <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>N</mml:mi></mml:math>-band Hopf insulator
- Max-value Entropy Search for Multi-Objective Bayesian Optimization.
- Probability Distribution of Hypervolume Improvement in Bi-objective Bayesian Optimization