Computational searches for new two-dimensional materials, meaning crystals one or a few atoms thick that can be peeled off a layered bulk parent, all start the same way. They compute an exfoliation energy, the energy cost of pulling one layer off the bulk. They discard every compound above a fixed cutoff, then rank the survivors by some useful property such as band gap or magnetism. The cutoff is an empirical calibration, not a law of physics. One major study uses 150 meV/atom, another uses 30–35 meV/Ų, and the right value depends on whether the layer will be peeled with tape or sonicated in solvent. A compound just over the line is thrown away no matter how good its electronic properties are. Changing the cutoff can also reorder the "best" candidates, and the uncertainty in the computed energies is comparable to the margins that decide the cut.
This note proposes to stop treating the cutoff as an input. The set of answers to "what is the best material whose exfoliation energy is below τ?", collected over all possible τ, is exactly the Pareto front between exfoliation energy and the target property. The Pareto front is the set of compounds that no other compound beats on both counts at once. So the scientifically useful object is that front, and the computational question is how few expensive calculations it takes to identify which compounds are on it. We propose an active-learning loop that spends density-functional theory (DFT) calculations, the standard quantum-mechanical energy method, only where they change our belief about front membership. This follows the logic that convex-hull-aware active learning applied to thermodynamic stability. We benchmark the loop retrospectively against the existing high-throughput exfoliation databases, where the true front is already known.
Core Claim
For two objectives, minimizing exfoliation energy $E_{exf}$ and maximizing a functional property $P$, the threshold-constrained winners $x^*(\tau) = \arg\max_x \{P(x) : E_{exf}(x) \le \tau\}$ swept over all $\tau$ trace out exactly the non-dominated staircase of the candidate pool. Adding thermodynamic stability (energy above the convex hull, $E_{hull}$) as a third objective gives the analogous three-dimensional front. Because of this identity, a pipeline that identifies the front answers every practitioner's threshold at once, and no single threshold has to be fixed in advance.
Our claim is that front membership is a global property of the pool, just as hull membership is. It should therefore be learned with an acquisition rule that targets membership uncertainty. Rules that target per-compound uncertainty, hypervolume gain, or a weighted sum of objectives should do worse. We expect this to recover the threshold-indexed answer set with markedly fewer expensive calculations than a funnel with a softened cutoff, than hypervolume-improvement Bayesian optimization, or than variance sampling. If it does not, the result is still informative: it would say that the existing funnel is close to optimal and that the field's threshold choice matters less than assumed.
Research Program
Three prior lines of work set up the problem. The screening studies of Mounet et al. and Ashton et al., together with the 2DMatPedia work of Shen et al., motivate it. Mounet et al. computed binding energies for 5,619 layered compounds with two dispersion-corrected functionals. Ashton et al. identified 826 layered Materials Project compounds by bond topology and cut at 150 meV/atom. Shen et al. built a database of more than 6,000 2D materials using an $E_{hull} < 150$ meV/atom stability filter. All three apply sequential hard cutoffs, so the 146 Ashton compounds above the line and Mounet's 789 "potentially exfoliable" compounds never compete on function.
Convex hull-aware active learning (CAL) supplies the method we adapt. CAL draws joint Gaussian-process posterior samples of phase energies, forms a probabilistic convex hull from each sample, and queries where the information gain about the hull is highest. It reports that plain variance sampling needed roughly twice as many calculations to resolve the same boundary. CAL handled only bulk energy hulls on synthetic binary and ternary systems. We replace the energy hull with the multi-objective non-dominated staircase.
BIRDSHOT is the closest existing multi-objective materials loop. It runs batch Bayesian optimization with expected hypervolume improvement (EHVI) on bulk alloys. We use it as our main competitor, not as a building block, because EHVI rewards finding good points and does not reward certainty about which points are good.
The acquisition rule works as follows. Fit independent Gaussian-process surrogates for $E_{exf}$, $P$ and $E_{hull}$, with inputs given by crystal embeddings from ALIGNN-FF or Crystalformer. Draw $S$ joint posterior samples of all objectives over the pool. For each compound $x$, let $p_x$ be the fraction of samples in which $x$ is non-dominated. The rule picks the compound $z$ whose calculation, with outcome $y_z$, most reduces the total membership entropy
$$a(z) = \sum_x h(p_x) - \mathbb{E}_{y_z}\Big[\sum_x h(p_x \mid y_z)\Big],$$
where $h(p) = -p\log p - (1-p)\log(1-p)$ is binary entropy. The expectation over $y_z$ is estimated by fantasizing a few posterior outcomes and conditioning on each.
Batching follows BIRDSHOT's batch sizes of 8 to 16, using greedy fantasies. Two cheap physics features enter as surrogate inputs. The first is the vdW-DF2-C09 versus rVV10 disagreement, where both values exist. The second is Mounet's "discrepancy trick," the interlayer expansion when dispersion forces are switched off. These matter because, as the file notes, the Gaussian distance decay in Crystalformer does not represent the long-range dispersion forces that set interlayer binding.
One analysis comes before any loop is run. We compute what fraction of the true front, in the existing data, is "supported," meaning reachable by some positive weighted sum. On a finite discrete pool, staircase points inside the convex hull of the front are unreachable by any linear weighting. This is the finite-set version of the concavity argument in Zhang's analysis of hypervolume scalarization. That fraction tells us immediately how much weighted-sum ranking can possibly recover.
Baselines
The naive reference point is the published funnel. It applies the surrogate-predicted exfoliation cutoff (150 meV/atom on the Ashton pool, 35 meV/Ų on the Mounet pool), then runs DFT on survivors in order of predicted property. A random-order variant serves as the floor.
The strongest simple competitor is a softened funnel, which we expect to be the real threat. It widens the cutoff by a margin $\delta$ to absorb surrogate error, and we tune $\delta$ on a held-out chemistry split under the same tuning budget as our hyperparameters. It then queries survivors in order of predicted property, or of an optimistic estimate of predicted property.
The strongest existing method is EHVI batch Bayesian optimization in the BIRDSHOT configuration, run on identical surrogates, identical batch sizes and identical expensive-call budgets. No reported result was found on Lacuna for any multi-objective method on exfoliation data. The only reported BIRDSHOT figure is on alloys: 45 Pareto points found while exploring 0.15% of a roughly 53,000-composition space. That number is not transferable, so we will rerun EHVI ourselves rather than quote it.
We also run hypervolume scalarization with random weights, following Zhang. It achieves provably optimal $O(T^{-1/k})$ hypervolume regret, but the reported BBOB comparison is qualitative only, and no number was found on Lacuna. Plain variance sampling, CAL's "BASE" baseline, completes the comparison.
Two ablations isolate the key ingredient. The first swaps our membership-entropy rule for EHVI while keeping everything else fixed. The second applies CAL-style hull entropy to $E_{hull}$ alone and ignores $E_{exf}$ and $P$, which tests whether the multi-objective target is what helps.
Our method beats the field only if it reaches 90% recall of true front members, with precision of at least 0.9, using fewer expensive calls than the best-tuned softened funnel and than EHVI. The margin must exceed seed-to-seed variation across 20 random initial designs.
Why not just…?
Why not just run DFT on every layered compound? For Mounet's 5,619 compounds at the standard dispersion-corrected level, this has already been done. A retrospective benchmark at that level therefore only simulates the regime that matters. The regime that matters is when the expensive objective is something not yet computed for the whole pool. Examples include energies corrected for the $d$- and $f$-electron errors that both Mounet et al. and Shen et al. flag, full-Brillouin-zone phonons in place of the zone-centre check Mounet et al. used, or Curie temperatures. We will state the speedup as a fraction of the pool, so readers can convert it to cost for whichever expensive calculation they care about. If the speedup is small, exhaustive calculation is the right answer, and we will say so.
Why not just widen the cutoff and rank by property? This is the most plausible threat. If surrogate errors are roughly uniform, a well-chosen margin may capture most front members at low cost. It will fail in two cases: where errors depend on chemistry, as with correlated transition-metal and lanthanide compounds, and where high-$P$ compounds sit deep above the cutoff. Our membership rule spends calls on those compounds selectively. The softened funnel is tuned and given equal budget, so the recall-versus-calls curves settle the question directly. We will also report the case where it wins.
Why not just use a weighted sum, or a constrained optimizer with a Lagrange multiplier? The supported-fraction analysis puts a hard cap on what any weighted sum can recover, and we compute that cap from the data before any loop runs. A constrained formulation, which in its reinforcement-learning version Paternain et al. show has no duality gap in the space of state-visitation distributions, answers one threshold per solve. Our identity says the front answers all thresholds at once. We test this by measuring per-threshold regret across a sweep of cutoffs.
Why not just use EHVI as BIRDSHOT did? EHVI may well match our rule early in a run. We expect it to waste late calls refining the extremes of the front while near-front compounds remain ambiguous. The EHVI-swap ablation measures exactly this difference.
Proposed Experiments and Decisive Checks
The first experiment is the retrospective study. The Mounet pool supplies $E_{exf}$ as binding energy in meV/Ų, using both functionals where available. The Ashton pool supplies 826 compounds with $E_{exf}$ in meV/atom. Band gap and magnetic moment from the released data, or from Materials Project and 2DMatPedia, serve as $P$, and Materials Project gives $E_{hull}$. We treat the stored DFT values as the expensive oracle and the ALIGNN-FF or Crystalformer predictions as the cheap prior. ALIGNN-FF's reported energy error is 0.086 eV/atom. That error is of the same order as the 150 meV/atom cutoff, though partial cancellation in energy differences may shrink it, and we will measure this directly.
The metrics are front recall and precision against the number of calls, hypervolume of the identified set, and threshold regret. Threshold regret is the gap between $P$ of the method's best identified compound below $\tau$ and $P$ of the true best compound below $\tau$, averaged over $\tau$ from 20 to 50 meV/Ų and from 100 to 200 meV/atom.
The second check measures front sensitivity to the choice of functional: what fraction of front members change status between vdW-DF2-C09 and rVV10. If the fraction is large, the expensive oracle itself is noisy, and the loop must model functional disagreement as observation noise.
The third experiment is prospective. We run the loop on Ashton's 146 above-cutoff compounds and Mounet's "potentially exfoliable" set, with full phonon dispersion as the expensive stability objective. Success means finding even a handful of dynamically stable compounds that dominate some funnel survivor on function. Such compounds are materials the field has discarded, and they would be the concrete payoff of the proposal.
References
- Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds
- Topology-Scaling Identification of Layered Solids and Stable Exfoliated 2D Materials
- High-throughput Discovery and Intelligent Design of 2D Functional Materials for Various Applications
- Probabilistic prediction of material stability: integrating convex hulls into active learning
- An Interoperable Multi Objective Batch Bayesian Optimization Framework for High Throughput Materials Discovery
- Unified Graph Neural Network Force-field for the Periodic Table
- Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding
- Optimal Scalarizations for Sublinear Hypervolume Regret
- Constrained Reinforcement Learning Has Zero Duality Gap