Accelerated Dopant Screening in Oxide Semiconductors via Multi-Fidelity Contextual Bandits and a Three-Tier DFT Validation Funnel

This paper introduces a multi-fidelity contextual bandit strategy combined with a three-tier DFT validation funnel to efficiently screen oxide semiconductor dopants, successfully identifying optimal Cu-containing co-doped ZnO systems for visible-light applications while reducing computational costs by 81% and revealing that dopant performance is governed by just two latent chemical dimensions.

The Gist

Imagine you are a master chef trying to invent the perfect new soup. You have a pantry with 5 different types of broth (the oxide hosts) and a spice rack with 16 different spices (the dopants). You want to mix them in every possible combination—single spices, double spices, triple spices—to find the one mix that tastes just right (a specific "band gap" for light absorption).

The problem? There are 529 possible recipes.

In the world of materials science, "tasting" a soup isn't as simple as taking a spoonful. To know if a recipe works, you have to run a massive, expensive, 4-hour simulation on a supercomputer. If you tried to taste every single one of the 529 recipes, it would take years and cost a fortune.

This paper is about a smart, fast, and safe way to find the best recipe without tasting everything.

Here is how the authors did it, broken down into three simple concepts:

1. The "Smart Taster" (Contextual Bandits)

Instead of tasting recipes randomly (like a blindfolded chef) or trying to guess the flavor of every possible soup at once (which is too slow), the authors used an algorithm called MF-OFUL.

Think of this algorithm as a super-smart sous-chef.

• The Strategy: The sous-chef looks at the ingredients you have (ionic size, charge, etc.) and makes a guess: "Based on what we've tasted so far, this new mix of Copper and Yttrium in Zinc broth is likely to be delicious."
• The Trick: The sous-chef doesn't just guess; it also knows when it's unsure. If it's unsure, it says, "Let's actually cook and taste this one." If it's very confident, it says, "I'm 90% sure this one will be good, so let's just trust my prediction and move on."
• The Result: By trusting its own "cheap" predictions 81% of the time, the team only had to run the expensive 4-hour computer simulations for about 20% of the recipes. They found the best soup in record time.

2. The "Three-Stage Safety Net" (The Validation Funnel)

Here is the catch: Sometimes, a computer simulation can lie. It might say a soup is delicious when it's actually salty, or vice versa. The authors realized that one type of computer test isn't enough to catch every mistake.

So, they built a Three-Tier Safety Net:

• Tier 1 (The Quick Sniff): A fast, rough simulation. It screens out the obvious failures.
• Tier 2 (The "Spice Check"): Some spices (like transition metals) have tricky electronic properties that the first test misses. This tier adds a specific correction (like adding a pinch of salt to balance the flavor) to catch those specific errors. Analogy: It's like realizing your "quick sniff" missed that the soup is actually too spicy because of a hidden ingredient.
• Tier 3 (The Texture Check): Sometimes, the ingredients don't fit together physically. The atoms might be too big for the space, causing the soup to "collapse." This tier checks if the physical structure holds up.

Why this matters: The authors found that some recipes looked great in Tier 1 but were actually terrible in Tier 2 or 3. Without this safety net, they might have wasted time on a "fake" winner.

3. The "Recipe Transfer" (Collaborative Filtering)

Imagine you are opening a new restaurant in a city you've never visited. You don't know what the locals like yet. This is the "Cold Start" problem.

The authors used a trick borrowed from Netflix recommendations.

• Netflix knows you like Action movies because you liked other Action movies.
• The authors realized that if a spice works well in Zinc broth, it's likely to work well in Magnesium broth because they are chemically "cousins."
• By looking at what worked in the broths they had already tested, they could guess what would work in the new ones immediately. This saved them from wasting time on random guesses at the very beginning.

The Big Discovery

Using this smart system, they found a "Golden Recipe": Zinc Oxide doped with Copper and Yttrium.

• Why it's cool: Most oxide semiconductors only absorb invisible UV light. This new mix absorbs visible light (the kind we can see).
• The Application: This is a huge step forward for solar power and water splitting. If we can make materials that absorb visible light efficiently, we can create better solar panels and cleaner fuels using sunlight.

Summary

The authors didn't just find a new material; they built a new way to discover materials.

1. They used a smart algorithm to skip 81% of the expensive testing.
2. They used a three-step safety check to ensure the results weren't fake.
3. They used recommendation logic to jump-start the search in new areas.

It's like turning a process that used to take a lifetime of blind tasting into a 2-day sprint where you only taste the most promising dishes, ensuring you find the perfect recipe without burning the kitchen down.

Technical Summary

1. Problem Statement

The bandgap engineering of oxide semiconductors (e.g., ZnO, TiO₂, SrTiO₃) via aliovalent doping is critical for photocatalysis and optoelectronics. However, the combinatorial space of dopant elements, substitution sites, concentrations, and co-doping pairs creates a search space of hundreds to thousands of candidates.

• The Bottleneck: Density Functional Theory (DFT) is the standard for predicting electronic properties but is computationally expensive. A single calculation can take minutes to days, making exhaustive screening of large candidate pools (e.g., >500 candidates) prohibitive.
• The Challenge: Existing Bayesian Optimization (BO) methods using Gaussian Processes (GPs) face scalability issues ($O(n^3)$ complexity) and sensitivity to kernel selection. Furthermore, standard DFT levels (like PBE) often fail to accurately predict properties for transition metal dopants due to self-interaction errors (d-electron delocalization) or geometric distortions, leading to qualitative misclassifications.

2. Methodology

The authors propose a holistic framework combining Multi-Fidelity Contextual Bandits, a Three-Tier DFT Validation Funnel, and Cross-Host Collaborative Filtering.

A. Multi-Fidelity Contextual Bandits (MF-OFUL)

Instead of standard BO, the authors employ OFUL (Optimism in the Face of Uncertainty for Linear bandits), a sequential decision algorithm.

• Linear Model: It fits a linear model relating compositional features (ionic radius, electronegativity, d-electron count, etc.) to bandgap rewards.
• Multi-Fidelity Strategy: To reduce DFT costs, the algorithm uses a Ridge regression surrogate as a "low-fidelity" evaluator.
  • Decision Logic: The algorithm selects candidates based on an Upper Confidence Bound (UCB). If the uncertainty (UCB width) is high, it triggers an expensive DFT calculation (high fidelity). If uncertainty is low, it uses the cheap surrogate prediction (low fidelity).
  • Stability: A Lyapunov stability analysis (formally verified in Lean 4) proves that surrogate errors do not accumulate and destabilize the optimization, provided specific error thresholds are met.
• Efficiency: This approach reduces computational complexity from $O(n^3)$ (in GP-based BO) to $O(d^2)$ per step, where $d$ is the feature dimension.

B. Three-Tier DFT Validation Funnel

Recognizing that no single DFT level is sufficient, the authors implement a tiered validation protocol:

1. Tier 1 (Fast Screening): PBE-SCF calculations guided by the bandit for initial ranking.
2. Tier 2 (d-electron Correction): PBE+U calculations for candidates with transition metal dopants ($d$-electrons > 0) to correct for on-site Coulomb repulsion and localization errors.
3. Tier 3 (Geometry Relaxation): Ionic relaxation for candidates with large ionic radius mismatches (>20%) to correct for geometric distortions that static calculations miss.

• Orthogonality: The tiers catch distinct failure modes: Tier 2 fixes electronic errors (e.g., V-doped ZnO), while Tier 3 fixes geometric errors (e.g., In-doped SrTiO₃).

C. Cross-Host Collaborative Filtering

To solve the "cold-start" problem when screening a new host with no data, the authors apply Matrix Factorization (SVD) to a host-dopant reward matrix.

• This reveals that dopant performance is governed by just two latent chemical dimensions, allowing the system to predict rankings in unseen hosts based on data from previously screened hosts.

D. Clinical Trial Adaptation

The framework adapts the RECOVERY trial design, treating dopant families (e.g., Rare Earths, 3d Transition Metals) as "treatment arms." An adaptive platform design drops futile families early and reallocates resources to promising ones, identifying Rare Earths as the most productive family without prior chemical knowledge.

3. Key Contributions

1. Algorithmic Innovation: Introduction of MF-OFUL, a multi-fidelity contextual bandit that replaces 81% of DFT evaluations with surrogate predictions while maintaining theoretical regret guarantees.
2. Validation Protocol: A Three-Tier Funnel that systematically catches orthogonal failure modes (electronic vs. geometric) that single-level DFT misses.
3. Formal Verification: A Lyapunov stability proof (1,826 lines in Lean 4) guaranteeing that the multi-fidelity switching mechanism remains stable and does not accumulate error.
4. Cross-Domain Transfer: Successful application of clinical trial adaptive designs and collaborative filtering (from recommendation systems) to materials discovery.
5. Open Benchmark: Release of a dataset containing 583 DFT calculations, screening code, and stability proofs.

4. Key Results

Screening Performance

• Efficiency: In a 529-candidate ZnO co-doping campaign, MF-OFUL found the global optimum in 100% of 50 independent trials using only 42 DFT calls (plus 184 surrogate calls).
• Comparison: It achieved 10x lower simple regret than standard Random screening and significantly outperformed OFUL (without surrogates) and Thompson Sampling.
• Cost Reduction: A closed-loop campaign that would have required ~440 CPU-hours was reduced to 62 CPU-hours.

Material Discovery

• Optimal Candidate: Y₂Cu₂ co-doped ZnO was identified as the optimal candidate with a bandgap of 1.84 eV (PBE) / 1.71 eV (HSE06), falling within the visible-light range (1.0–1.8 eV).
• Systematic Trends: Cu-containing co-doped systems consistently achieved visible-light bandgaps.
• Validation Insights:
  • V-doped ZnO: PBE predicted it as near-metallic, but PBE+U corrected it to a wide-gap insulator (+3.69 eV shift).
  • Cu-doped SrTiO₃: PBE predicted a wide gap, but PBE+U revealed it enters the visible-light window (1.59 eV).
  • SrTiO₃:In: PBE predicted a metallic state, but ionic relaxation (Tier 3) corrected it to a wide gap (3.48 eV) by removing spurious midgap states caused by lattice distortion.

Cross-Host Transfer

• The host-dopant interaction matrix was found to be low-rank (2 components explain ~97% of variance).
• Using Collaborative Filtering for warm-starting improved discovery rates by 53% in new hosts compared to cold-start bandits.

5. Significance

This work represents a paradigm shift in computational materials discovery:

• Scalability: It demonstrates that linear bandits with multi-fidelity surrogates can scale to large candidate pools ($>500$) where Gaussian Process-based BO fails due to computational cost.
• Reliability: By integrating a multi-tier validation funnel, the method ensures that the "best" candidate found is not an artifact of DFT approximations (e.g., self-interaction error).
• Reproducibility & Rigor: The formal verification of stability in Lean 4 provides a level of mathematical rigor rarely seen in applied materials informatics, ensuring the algorithm is safe to deploy.
• Generalizability: The framework is applicable to any oxide host and can be extended to larger pools (e.g., ternary co-doping) by adding lower-fidelity tiers (e.g., Neural Network Potentials).

The paper concludes that the integration of MF-OFUL (for efficient search), the Three-Tier Funnel (for accurate validation), and Collaborative Filtering (for knowledge transfer) creates a complete, formally grounded protocol for high-throughput dopant discovery.