Machine-learning Guided Search for Phonon-mediated Superconductivity in Boron and Carbon Compounds

This study presents a machine-learning-guided workflow that combines *ab-initio* calculations with the inclusion of dynamically unstable compounds to identify promising boron and carbon-based superconductors, predicting several new candidates with critical temperatures ranging from 4.0 K to 35 K.

The Gist

Imagine you are a chef trying to invent the perfect new dish: a material that can conduct electricity with zero resistance (superconductivity) without needing to be frozen in liquid helium. You know that some ingredients, like Boron and Carbon, are great at this, but finding the right recipe among thousands of possibilities is like searching for a needle in a haystack.

This paper describes a new, smart way to find that needle. The authors, Niraj K. Nepal and Lin-Lin Wang, built a "robot chef" (Machine Learning) that works alongside a "super-precise taste tester" (Quantum Physics calculations) to discover new superconducting recipes.

Here is the story of their discovery, broken down into simple concepts:

1. The Problem: The "Broken" Ingredients

Usually, when scientists look for new materials, they throw away any recipe that looks "broken" or unstable. In the world of atoms, a "broken" recipe means the atoms are wiggling in a way that suggests the structure would collapse.

• The Old Way: "This recipe is unstable? Throw it in the trash! We only want perfect, stable structures."
• The New Insight: The authors realized that sometimes, these "wobbly" structures are actually hiding a secret superpower. If you give them a little push (like applying pressure or tweaking the electrons), they stop wobbling and become super-strong superconductors. They decided to stop throwing away the "broken" recipes and start investigating them.

2. The Workflow: The Smart Search Party

They created a three-step loop to find the best materials:

• Step 1: The Big Library (Data Extraction): They started with a massive digital library of about 1,500 Boron and Carbon compounds. They filtered out the ones that were magnetic or too big, leaving them with about 700 candidates.
• Step 2: The Taste Test (DFPT Calculations): They used a powerful computer method called Density Functional Perturbation Theory (DFPT) to simulate how these materials behave. This is like running a high-tech simulation to see if the material conducts electricity well.
  • The Hurdle: Sometimes the computer simulation gets "confused" because the grid of atoms it's looking at is too coarse. The authors invented a clever "convergence test" (a mathematical checklist) to make sure the computer wasn't just guessing.
• Step 3: The Robot Chef (Machine Learning): This is the magic sauce. They trained two different AI models (called CGCNN and ALIGNN) on the data from the taste tests.
  • The Twist: Most previous AI models were only trained on "perfect" materials. These authors trained their AI on both perfect materials and the "wobbly" (unstable) ones they had stabilized.
  • The Result: The AI model named ALIGNN turned out to be the better chef. It was especially good at understanding the "wobbly" materials because it paid attention to the angles between atoms, not just the distances.

3. The Discovery: Finding the Gold

By using this loop, they predicted several new superconductors. Here are the highlights:

• The "Almost Perfect" Ones: They found stable materials like TaNbC2 (Tantalum-Niobium-Carbon) which could superconduct at 28.4 Kelvin. That's very cold, but much warmer than absolute zero, and it's a huge step up from what we have now.
• The "Wobbly" Ones (The Big Surprise): They found that materials with "wobbly" atoms, once stabilized, could get even hotter.
  • Ca5B3N6: This compound, which has a cage-like structure, could superconduct at a scorching 35 to 42 Kelvin.
  • MoRuB2 and RuVB2: These Ruthenium-based compounds also showed promise, reaching around 15 Kelvin.

4. Why This Matters

Think of the search for superconductors like looking for a new planet.

• Before: Astronomers only looked at planets that were already stable and orbiting perfectly. They missed the ones that were wobbling in their orbits but might have been habitable if nudged slightly.
• Now: This paper says, "Let's look at the wobbling planets too!"

By including these "unstable" compounds in their training data, the AI learned a much richer picture of how materials work. It proved that completeness is key: an AI is only as good as the data you feed it. If you leave out the "weird" data, the AI misses the best discoveries.

The Bottom Line

The authors didn't just find a few new materials; they changed the rules of the game. They showed that by combining smart computer simulations with a "don't throw anything away" attitude toward unstable structures, we can find new superconductors that might one day help us build lossless power grids, faster trains, or even quantum computers.

In short: They built a smart robot that learned to cook with "broken" ingredients, and it just served up some of the hottest (in a good way) superconductors we've ever seen.

Technical Summary

1. Problem Statement

The search for high-temperature superconductors (SC) at ambient pressure remains a significant challenge in condensed matter physics. While high-pressure hydrides (e.g., $H_3S$) have shown near-room-temperature superconductivity, they require extreme pressures. Boron (B) and Carbon (C) compounds are promising candidates due to their light atomic mass and strong electron-phonon coupling (EPC), yet finding new materials with critical temperatures ($T_c$) significantly higher than the benchmark $MgB_2$ (39 K) has proven difficult.

Two major bottlenecks hinder the discovery of new B/C-based superconductors via computational methods:

1. Computational Convergence: Accurate calculation of EPC properties using Density Functional Perturbation Theory (DFPT) requires dense Brillouin zone (BZ) sampling ($k$- and $q$-meshes), which is computationally expensive. Coarse meshes often lead to inaccurate $T_c$ predictions.
2. Dynamic Instability: Traditional high-throughput screening discards compounds with imaginary phonon modes (dynamic instability). However, recent evidence suggests that stabilizing these modes can yield significant EPC contributions and high $T_c$, a factor largely overlooked in previous machine learning (ML) studies.

2. Methodology

The authors developed an iterative, machine-learning-guided workflow that combines ab-initio calculations with ML models to screen 417 Boron, Carbon, and Borocarbide compounds.

A. Data Extraction and Filtering

• Source: Materials Project (MP) database.
• Criteria: Metallic, negative formation energy, non-magnetic, excluding oxides, $C_{60}$, and most lanthanides.
• Constraints: Primitive cells with $\le 40$ atoms and $\le 4$ element types.
• Initial Pool: Reduced from ~1500 to ~700 candidates (121 known $T_c$, 579 unknown $T_c$).

B. Ab-initio Calculations (DFPT)

• Method: Density Functional Perturbation Theory (DFPT) with the isotropic Eliashberg approximation.
• Convergence Strategy (Ansatz Test): To address the high cost of dense mesh sampling, the authors developed a convergence test based on the decay of $T_c$ with respect to Gaussian broadening ($\sigma$).
  • They modeled $T_c$ decay as $T_c = \exp(-A\sigma^{1/3} + B)$.
  • Using $MgB_2$ as a reference, they established a threshold decay parameter $A_{MgB_2} \approx 12-13$.
  • If a compound's calculated $A > 12$, the mesh is deemed unconverged, and calculations are repeated with denser grids.
• Handling Instability: For dynamically unstable compounds (imaginary phonon modes), the authors employed three stabilization techniques:
  1. Lattice Distortion: Relaxing the structure along the imaginary eigenmode.
  2. Pressure: Applying hydrostatic pressure (5–60 GPa).
  3. Electronic Smearing: Increasing electronic smearing (0.05–0.1 Ry) to stabilize modes without full structural relaxation (crucial for high-throughput efficiency).

C. Machine Learning Models

• Models: Two Graph Neural Networks (GNNs) were trained:
  1. CGCNN: Crystal Graph Convolutional Neural Network.
  2. ALIGNN: Atomistic Line Graph Neural Network (incorporates bond angles).
• Workflow:
  • Run 1: Trained on 250 dynamically stable compounds (109 SC, 141 NSC).
  • Run 2: Used Run 1 models to predict $T_c$ for remaining candidates, selected top candidates for DFPT, and expanded the dataset to 323 compounds.
  • Run 3: Incorporated 28 stabilized dynamically unstable compounds into the training set (total 351 training samples) to test model performance on unstable systems.

3. Key Contributions

1. Inclusion of Dynamically Unstable Compounds: This is the first study to systematically include compounds with imaginary phonon modes in the training dataset for phonon-mediated superconductivity. These compounds constitute ~20% of the studied dataset.
2. Convergence Ansatz Test: The development of a robust, low-cost test ($A$-parameter) to identify and correct non-converged DFPT results, significantly improving the reliability of the training data.
3. Model Architecture Comparison: The study demonstrates that ALIGNN consistently outperforms CGCNN, particularly when the dataset includes dynamically unstable compounds. The inclusion of bond angles in ALIGNN is critical for capturing the physics of soft phonon modes and lattice distortions.
4. Stabilization Protocols: Demonstrated that electronic smearing is an effective proxy for pressure in stabilizing imaginary phonon modes for high-throughput screening, yielding results comparable to full pressure calculations.

4. Key Results

• Model Performance:
  • ALIGNN achieved lower Mean Absolute Errors (MAE) than CGCNN across all properties ($\lambda$, $\omega_{log}$, $T_c$).
  • For the independent test set including unstable compounds, ALIGNN achieved an MAE of 4.3 K for $T_c$ and 3.5 K for post-processed $T'_c$, compared to 4.5 K and 5.3 K for CGCNN.
  • The models performed best in predicting $\omega_{log}$ (directly related to bonding strength) compared to $\lambda$ and $T_c$.
• Predicted Superconductors:
  The workflow identified several promising candidates with formation energies near the ground state convex hull:
  • Dynamically Stable:
    • $TaNbC_2$: $T_c \approx 28.4$ K ($\Delta E_h \approx 0.02$ eV/atom).
    • $Nb_3B_3C$: $T_c \approx 16.4$ K.
    • $Y_2B_3C_2$: $T_c \approx 4.0$ K.
  • Dynamically Unstable (Stabilized):
    • $Ca_5B_3N_6$: $T_c \approx 35$ K (up to 42.4 K with lower Coulomb pseudopotential). Stabilized via electronic smearing; exhibits significant EPC from soft modes at the H-point.
    • $MoRuB_2$: $T_c \approx 15.6$ K.
    • $RuVB_2$: $T_c \approx 15.0$ K.
    • $Pd_3CaB$: $T_c \approx 7.0$ K.
    • $RuSc_3C_4$: $T_c \approx 6.6$ K.
• Validation: The calculated $T_c$ for known compounds (e.g., $MgB_2$, $Y_2C_3$) showed good agreement with experimental data after applying the convergence ansatz and stabilization techniques.

5. Significance

• Bridging the Gap: The study addresses the large materials gap between $MgB_2$ (39 K) and high-pressure hydrides, identifying potential ambient-pressure superconductors with $T_c$ values competitive with or exceeding current known limits for B/C systems.
• Paradigm Shift in Screening: By proving that dynamically unstable compounds are not just noise but potential high-$T_c$ candidates, the paper advocates for a fundamental shift in high-throughput screening protocols to include stabilization techniques rather than discarding unstable structures.
• ML Best Practices: The work highlights that the completeness of the training dataset (including diverse stability regimes) is as crucial as the model architecture itself. It validates ALIGNN as a superior tool for materials discovery involving complex lattice dynamics.
• Experimental Guidance: The specific predictions (e.g., $Ca_5B_3N_6$, $TaNbC_2$) provide concrete targets for experimental synthesis and verification, potentially accelerating the discovery of new ambient-pressure superconductors.