Active-Learning Inspired $\textit{Ab Initio}$ Theory-Experiment Loop Approach for Management of Material Defects: Application to Superconducting Qubits

1. Problem Statement

Superconducting quantum computing devices, particularly those based on niobium (Nb) Josephson junctions (JJs), suffer from performance degradation due to two-level systems (TLSs). These TLSs are often associated with surface oxides (specifically niobium oxide) that form at interfaces. These defects act as parasitic resonators, coupling to the qubit's electric field and significantly reducing coherence times.

While strategies like high-temperature heat treatment and etching exist, they are often empirical. A more promising approach is encapsulation with an oxygen diffusion barrier (a metal capping layer) to prevent oxygen from reaching the niobium substrate. However, selecting the optimal barrier material has historically been a trial-and-error process with limited predictive guidance, especially in data-sparse regimes where experimental synthesis is slow and expensive.

2. Methodology

The authors propose a closed-loop, active-learning framework that integrates first-principles theory, machine learning, and targeted experiments. The workflow consists of the following steps:

• First-Principles Calculations (DFT):

  • Density Functional Theory (DFT) calculations were performed using the JDFTx code.
  • Descriptors: Instead of relying on complex kinetic models, the study focuses on thermodynamic descriptors governing oxygen diffusion:
    1. Oxygen Interstitial Energy ($E(O_I)$): Energy to insert an oxygen atom into the metal capping layer.
    2. Oxygen Vacancy Energy ($E(V_O)$): Energy to remove an oxygen atom from the metal oxide.
  • The study calculated these energies for a wide range of candidate metals (noble metals, refractory transition metals, etc.).
  • Key Finding: Diffusion barrier heights (kinetics) showed no correlation with experimental oxide suppression, confirming that thermodynamics controls the outcome.

• Machine Learning Model:

  • A Logistic Regression model was employed to predict the probability ($P$) of niobium oxide formation.
  • Input: The defect energies ($E(O_I)$ and $E(V_O)$) serve as features.
  • Training Data: A limited set of experimental outcomes (XPS measurements) where metals were classified as "success" (oxide suppressed) or "failure" (oxide formed).
  • Handling Uncertainty: The model uses "soft labels" (probabilities between 0 and 1) rather than binary labels to account for experimental noise and fabrication defects.

• Active Learning Loop:

  1. The model predicts the likelihood of success for untested candidates based on initial DFT data.
  2. High-priority candidates are synthesized and tested experimentally (XPS).
  3. New experimental data is fed back to retrain the model, refining the decision boundary.
  4. The cycle repeats to identify optimal materials with minimal experimental iterations.

3. Key Contributions

• Identification of a Single Dominant Descriptor:
  The study reveals that the complex two-dimensional descriptor space (Interstitial vs. Vacancy energy) can be reduced to a single, physically interpretable metric: Oxide Formation Energy per Oxygen Atom ($E_{form}/m$).

  • The logistic regression decision boundary corresponds almost perfectly to a linear combination of the original descriptors, which maps directly to the thermodynamic cost of forming a metal oxide normalized by oxygen content.
  • Rule of Thumb: If forming the metal oxide is thermodynamically favorable (low/negative formation energy), the metal will effectively block oxygen from reaching the niobium.

• Integration of Structural Coherence:
  Beyond thermodynamics, the authors introduce lattice mismatch as a secondary criterion. Even if a metal is thermodynamically stable, a large lattice mismatch with Nb can create grain boundaries or dislocations that act as diffusion pathways.

  • This dual-criteria approach (Thermodynamics + Lattice Matching) refines the candidate list.

• Active-Learning Framework for Sparse Data:
  The paper demonstrates a successful implementation of an interpretable active-learning loop specifically designed for materials science problems where experimental data is scarce. It bridges the gap between high-throughput computational screening and low-throughput experimental validation.

4. Results

• Predicted Candidates:
  The model successfully identified Zirconium (Zr), Hafnium (Hf), and Tantalum (Ta) as effective diffusion barriers.

  • Zr emerged as the top candidate, satisfying all thermodynamic criteria and showing a chemically sharp interface with Nb.
  • Scandium (Sc) was identified as a highly promising candidate that had not yet been experimentally tested at the time of the study.

• Experimental Validation:

  • Successes: Experimental XPS measurements confirmed that Zr, Hf, and Ta successfully suppressed niobium oxide formation.
  • Failures: Metals like Aluminum (Al) and Titanium (Ti) showed higher lattice mismatches or less favorable thermodynamics, leading to less optimal performance or interface instability.
  • New Discoveries: The loop successfully predicted the behavior of Ti, Mo, and Ir in subsequent iterations.

• Descriptor Map Analysis:

  • The "Descriptor Map" (Oxygen Vacancy Energy vs. Metal Interstitial Energy) clearly separates successful barriers (upper-left quadrant) from failures (lower-right).
  • The decision boundary derived from the logistic regression aligns with the line of constant oxide formation energy per oxygen atom.

5. Significance

• Rational Design of Qubit Materials: This work provides a concrete, predictive pathway to solving the TLS problem in superconducting qubits, potentially leading to significant improvements in coherence times and device performance.
• Methodological Advancement: It establishes a blueprint for using interpretable machine learning in materials discovery. By reducing complex defect energetics to a single physical descriptor (oxide formation energy), the model offers physical insight rather than acting as a "black box."
• Efficiency: The approach minimizes the number of expensive experimental iterations required to find optimal materials, making it highly applicable to other material design challenges where synthesis and characterization are bottlenecks.
• Future Directions: The identification of Scandium (Sc) and the validation of Zr-based capping layers open new avenues for fabricating next-generation, high-coherence superconducting quantum circuits.

In conclusion, the paper successfully demonstrates that a closed-loop strategy combining DFT-calculated defect energetics, logistic regression, and targeted experimentation can rationally design oxidation-resistant interfaces for quantum devices, moving the field from empirical trial-and-error to predictive materials science.