Temporal interference stimulation tries to reach deep brain structures, such as the hippocampus or thalamus, from the scalp. It applies two kilohertz currents at slightly different frequencies. Neurons ignore the kilohertz carrier, but where the two fields overlap they respond to the slow "beat" between them (Challenges and opportunities of an emerging deep brain stimulation approach in research; The neuron mixer and its impact on human brain dynamics). Steering that beat precisely needs many electrodes and a montage computed for each person's head. Electric fields spread very differently through different skulls and folds of cortex, so a montage designed on a template brain does not reliably hit the target in an individual (MOVEA). The standard pipeline first builds a finite-element head model, meaning a numerical mesh of scalp, skull, fluid, grey and white matter. It then runs one expensive field simulation per candidate electrode position to get that electrode's "lead field," the field it would produce per unit current. Only after that does a cheap but non-convex search over montages begin. For a dense cap, the per-electrode solves are the dominant personalization cost. They are paid up front, in full, for electrodes the final montage will never use.
This note proposes treating each per-electrode finite-element solve the way efficient architecture search treats training a candidate network: as an expensive evaluation to be avoided unless it could change a decision. The search starts from a small set of solved electrodes. It scores unsolved electrode positions with a cheap interpolated proxy for their lead field. A feedback controller then decides when to buy a real solve: when the search stagnates, or when the proxy's uncertainty could flip which of two montages dominates. The claim is that a personalized Pareto front of deep-target intensity versus focality can be recovered, matching the full-cap front, with a small fraction of the per-electrode solves.
Core Claim
The obvious combination at this intersection would be to put a neural surrogate or pairwise-ranking classifier in front of montage evaluation, as SMEMNAS does for architectures. That attacks the wrong cost. Quasi-static fields are linear in the injected currents. Once lead fields $L_j(\mathbf r)$ exist for every electrode $j$, any candidate montage is just a weighted sum:
$$\mathbf E_k(\mathbf r) = \sum_j L_j(\mathbf r)\, I_{k,j}, \qquad k \in \{1,2\},$$
Here $I_{k,j}$ is the current that frequency channel $k$ injects through electrode $j$. Candidate evaluation becomes a matrix product that runs quickly on a GPU. The non-convexity MOVEA emphasizes comes from two sources. The first is discrete electrode selection. The second is the beat amplitude, which along an orientation $\mathbf n$ is
$$A_{\mathbf n}(\mathbf r) = \big|\, |(\mathbf E_1+\mathbf E_2)\cdot \mathbf n| - |(\mathbf E_1-\mathbf E_2)\cdot \mathbf n| \,\big|.$$
This quantity is nonlinear in the two fields. It is not the cost of evaluating a candidate.
What is expensive is building the matrix $L$ for a new person, at one full mesh solve per electrode. Focality requires whole-brain fields, so a reciprocity shortcut that only yields field values at a few target points does not remove this cost. Our claim is therefore specific. Adaptive acquisition of lead fields, guided by the search's own state, reaches the full-cap Pareto front with far fewer solves than computing all of them. It also does better than a fixed sparse electrode grid of the same solve budget.
Research Program
The method has three parts.
The proxy. For an unsolved scalp position, we predict its lead field by interpolating along the scalp surface from already-solved neighbors, using geodesic-distance weights. We give the prediction an error bar. That error bar is calibrated from leave-one-out interpolation error among solved electrodes in the same scalp region. This plays the role of EZNAS's zero-cost proxies, which score architectures at initialization without training. Unlike EZNAS, our proxy is a physical approximation with an uncertainty attached, not an evolved formula. The uncertainty is what makes it usable for decisions.
The search. We use a multi-objective evolutionary search over mixed discrete-continuous montages: which electrodes are active and what current each carries. The objectives are beat amplitude at the target, beat amplitude outside the target, and optional avoidance zones, as in MOVEA. We adopt the two-population structure of SMEMNAS: one population exploits the current front, the other preserves diversity by crowding distance. From Adaptive TPOT we take the "stump start." Search begins on a classical four-electrode, two-pair layout plus a coarse ring of solved positions. Complexity grows only when simpler montages plateau.
The controller. This is the adaptive-parameter-control component, and it governs the evaluation budget rather than only mutation rates. That gap is named in the file as "rigid evaluation budgets." The controller buys a real solve for an unsolved electrode under either of two conditions.
The first is a dominance-flip test. The electrode is used by a candidate on or near the current front, and its proxy error bar is wide enough to reverse a pairwise dominance comparison between that candidate and a rival. This is the same pairwise framing SMEMNAS uses, but applied to deciding when to evaluate exactly, not to replacing evaluation.
The second is a stagnation trigger. Front hypervolume, meaning the volume of objective space dominated by the current Pareto front, has not improved for a set number of generations. The controller then solves the most promising unsolved positions in under-explored scalp regions. This mirrors stagnation-keyed schedules (Targeted Adversarial Examples for Black Box Audio Systems) and the self-adjusting mechanisms analyzed in Editorial.
Mutation step size on current amplitudes follows population spread in objective space, as Adaptive TPOT does with fitness standard deviation. Because objectives conflict, spread is measured per objective. Adaptation fires only when both objectives' spreads collapse. This is our concrete answer to the multi-objective instability flagged for naive scalar rules.
The second-order hyperparameter problem is real. The controller has its own knobs: the flip-test confidence level, the stagnation window, and the per-trigger batch size. We report sensitivity across a grid of each and treat strong sensitivity as a failure of the method.
Baselines
The full-cap reference computes lead fields for every electrode in a dense cap and runs the same search with exact evaluation throughout. This is the "do nothing clever" point and the quality ceiling. Every other method is scored by the solves it needs to reach the reference front's hypervolume within a tolerance. It shares the identical head meshes, conductivities, objectives and search code.
The strongest existing method is MOVEA, a genetic algorithm followed by multi-objective particle swarm optimization on fully computed lead fields (MOVEA). That paper reports high-definition temporal interference reaching about 0.51 V/m at deep targets, comparable to high-definition alternating-current stimulation. It reduces activated volume outside the target by up to 60%. It holds 0.3 V/m in the thalamus while steering current away from primary motor cortex. We reimplement MOVEA with its published structure and give it the same evaluation budget and tuning budget as our search. Beating it means matching its front, including the thalamus-with-avoidance case, at a fraction of its lead-field solves. The file reports no runtime or solve counts for MOVEA, so the cost comparison will be measured by us, not quoted.
The fixed sparse grid baseline uses the same number of solves as our method spends, but on a uniform subset of cap positions chosen in advance. It isolates whether adaptivity, rather than sparsity alone, matters.
The gradient-based continuous baseline optimizes currents directly by gradient descent over the full lead-field matrix, with a smooth relaxation of electrode on/off choices in the spirit of Differentiable Neural Network Architecture Search. No reported result was found on Lacuna for gradient-based temporal interference montage optimization.
Two ablations isolate the key ingredients. The first keeps the proxy but acquires on a fixed schedule, solving the top neighbors every few generations. The second keeps adaptive acquisition but removes the proxy, so unsolved electrodes simply cannot be used.
Why not just…?
Why not just compute all lead fields once and be done? For a research cohort this is affordable, which is why it is our reference. It does not scale to a clinical workflow where every patient needs a new mesh and a dense cap, and the review literature names personalization cost as the barrier (Challenges and opportunities of an emerging deep brain stimulation approach in research). Our headline number is exactly the fraction of these solves we avoid at matched front quality. If the answer is "almost none," the idea fails cleanly.
Why not just use a standard coarse electrode layout? This is the most serious threat. If deep-target optima reliably sit near standard positions, a fixed sparse grid could match us. The equal-solve fixed-grid baseline tests this directly. We additionally report how far the best electrodes in the full-cap optimum lie from the nearest coarse-grid position, per subject and target. If that distance is consistently small, we will report that a well-chosen fixed grid suffices, and the contribution narrows to choosing that grid per target.
Why not just optimize currents by gradient descent instead of evolution? Gradient methods handle the continuous currents well once $L$ is complete, and may beat evolution on that sub-problem. That is why they appear as a baseline. They do not address which lead fields to compute. They also struggle with the beat amplitude's absolute values and with discrete electrode choice, the source of the non-convexity reported by MOVEA. If gradient descent proves better at current allocation, we will swap it into our inner loop. The lazy-acquisition controller is independent of the inner optimizer.
Why not just use a template head's lead fields? This costs zero solves. MOVEA reports large cross-subject divergence in optimal protocols, so we expect it to underperform. We include it to measure the gap rather than assume it.
Proposed Experiments
We will use structural T1/T2 scans from the Human Connectome Project, the dataset MOVEA uses, to build individual head meshes with a dense candidate cap. Targets are hippocampus, thalamus, and thalamus with a motor-cortex avoidance zone.
For each subject and target we record four outcomes. The first is the number of finite-element solves, and wall-clock time, to reach 95% of the full-cap reference hypervolume. The second is beat amplitude at the target at matched off-target volume. The third is the proxy's pairwise-dominance accuracy and Kendall rank correlation against exact evaluation on front candidates. For context only, EZNAS reports a correlation of 0.65 on NASBench-201 CIFAR-10, a different problem. The fourth is controller sensitivity across its three knobs.
We run three decisive checks. First, adaptive acquisition must beat the equal-budget fixed grid on most subjects; otherwise the core claim is false. Second, the proxy-flip trigger must beat the fixed-schedule ablation at equal solves, which shows that uncertainty-driven acquisition, not just more solves, matters. Third, results must hold under held-out subjects with controller settings frozen from a development set. This tests the claim that static settings fail to transfer across anatomies while ours do. A further check, in the spirit of Cell-specific effects of temporal interference stimulation on cortical function, reports the carrier-field magnitude in superficial cortex for every montage. Lazy acquisition must not trade solves for hidden off-target carrier exposure.
References
- Challenges and opportunities of an emerging deep brain stimulation approach in research
- The neuron mixer and its impact on human brain dynamics
- Multi-objective optimization via evolutionary algorithm (MOVEA) for high-definition transcranial electrical stimulation of the human brain
- A Pairwise Comparison Relation-Assisted Multiobjective Evolutionary Neural Architecture Search Method With Multipopulation Mechanism
- EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring
- An Adaptive and Near Parameter-free Evolutionary Computation Approach Towards True Automation in AutoML
- Targeted Adversarial Examples for Black Box Audio Systems
- Editorial
- Differentiable Neural Network Architecture Search
- Cell-specific effects of temporal interference stimulation on cortical function