First-Principles Optical Descriptors and Hybrid Classical-Quantum Classification of Er-Doped CaF$_2$

This study presents a physics-informed machine learning framework that utilizes first-principles optical descriptors derived from DFT and LR-TDDFT calculations to successfully discriminate between pristine and Er-doped CaF$_2$, demonstrating that hybrid quantum neural networks can achieve high classification accuracy comparable to classical baselines despite the noise constraints of current quantum hardware.

The Gist

Imagine you have two jars of glass. One jar is made of pure, perfect crystal (let's call it Pure Glass). The other jar is made of the same crystal, but someone has sprinkled a tiny bit of special "magic dust" into it (let's call it Dusted Glass).

To the naked eye, they look almost identical. But if you shine a specific kind of light through them, they react differently. The "magic dust" changes how the glass absorbs and reflects that light, creating a unique "fingerprint" that proves the dust is there.

This paper is a story about how scientists tried to teach computers to spot that difference using two different types of "brains": a Classical Brain (the kind we use in regular computers today) and a Quantum Brain (a futuristic, experimental kind of computer that uses the weird rules of quantum physics).

Here is how they did it, step-by-step:

1. The Setup: Building the Glass

First, the scientists didn't use real glass jars. They built tiny, digital models of them inside a supercomputer.

• The Pure Model: A cluster of Calcium and Fluorine atoms (CaF₂).
• The Dusted Model: The same cluster, but they swapped one Calcium atom for an Erbium atom (the "magic dust").
• The Test: They used a complex mathematical method (called DFT and TDDFT) to simulate what happens when light hits these models. They calculated how the light gets absorbed at different energy levels, creating a long list of numbers that describe the "optical fingerprint" of each jar.

2. Picking the Right Clues

The computer generated thousands of data points for each jar. It was like having a 10,000-page book describing the glass, but most of the pages were boring or repetitive.
The scientists needed to find the three most important sentences in the book that would tell them which jar was which. They used a smart filter to pick the "Top 3 Clues":

1. How much light is absorbed (Absorption coefficient).
2. How much light is lost or dimmed (Extinction coefficient).
3. The specific energy color of the light (Transition energy).

These three numbers became the "ID card" for each jar.

3. The Race: Classical vs. Quantum Brains

Now, they set up a competition to see which type of computer could best tell the Pure Glass from the Dusted Glass using only those three ID cards.

Contestant A: The Classical Brain (SVM)

This is a standard, powerful computer algorithm. It looked at the data and drew a line to separate the two groups.

• The Result: It was incredibly good. It got 98.3% of the answers right. It was like a master detective who never misses a clue.

Contestant B: The Quantum Brain (QSVM)

This is a new type of algorithm designed to run on quantum computers. It tries to find patterns in a "quantum space" that regular computers can't easily see.

• On a Perfect Simulator (No Noise): It got 85.1% right. Good, but not as good as the classical brain.
• On a Noisy Simulator (With Errors): It got 81.7% right. The "noise" (like static on a radio) made it slightly worse.
• On Real Hardware (The IBM Quantum Computer): They ran it on an actual quantum chip in the real world. Because real quantum computers are currently very sensitive to errors and "decoherence" (losing their quantum state), the score dropped to 73.3%. It was still better than random guessing (50%), but it struggled with the messy reality of the hardware.

Contestant C: The Hybrid Quantum Brain (QNN)

This was a different approach. Instead of just looking for a static pattern, this was a "learning" quantum circuit. It was like a student taking a test, getting feedback, and adjusting its thinking to get better.

• The Result: This one did surprisingly well! It achieved 93% accuracy. It learned to navigate the quantum space better than the static QSVM, getting much closer to the performance of the classical brain.

The Big Takeaway

The paper concludes with a few key lessons:

1. The "Magic Dust" Leaves a Trace: The Erbium atoms definitely change the way the material interacts with light. These changes are strong enough to be detected by computers.
2. Classical is Still King (For Now): The regular computer (Classical SVM) was the most accurate and reliable. It proved that for this specific task, we don't need quantum computers yet to get great results.
3. Quantum is Promising but Noisy: The quantum computers (especially the real hardware one) made mistakes because they are currently fragile and prone to errors. However, the "learning" quantum model (QNN) showed that if we can fix the hardware issues, quantum computers might eventually learn complex patterns that are hard for regular computers to find.
4. It's a Benchmark: This study isn't about building a new laser or a medical device right now. It's a "stress test" to see how well current quantum machines can handle scientific data compared to old-school methods.

In short: The scientists proved that you can use light to tell the difference between pure and doped crystals. They then tested if a futuristic quantum computer could do this better than a normal one. The normal computer won the race, but the quantum computer showed it has potential to catch up if we can make it less "noisy."

Technical Summary

Technical Summary: First-Principles Optical Descriptors and Hybrid Classical–Quantum Classification of Er-Doped CaF2

Problem Statement
The identification of rare-earth-doped materials, specifically distinguishing pristine calcium fluoride (CaF2) from Erbium-doped CaF2 (CaF2:Er), presents a challenge in materials informatics. While rare-earth ions like Er3+ are critical for photon upconversion applications, their optical signatures are complex, involving intermediate electronic states and sequential energy transfer pathways that are difficult to capture directly via standard Time-Dependent Density Functional Theory (TDDFT). Conventional ground-state DFT cannot describe optical excitations, and while linear-response TDDFT provides excitation energies, the resulting high-dimensional spectral data requires robust, interpretable descriptors to enable systematic comparison and classification. Furthermore, while Machine Learning (ML) has been applied to spectral analysis, the potential of Quantum Machine Learning (QML) to capture nontrivial relationships in quantum-mechanical data remains an open area of exploration, particularly regarding whether quantum feature embeddings offer advantages over classical baselines in the context of noisy intermediate-scale quantum (NISQ) hardware.

Methodology
The study employs a physics-informed workflow integrating first-principles calculations with classical and quantum machine learning:

1. First-Principles Calculations:

   • System Construction: Finite clusters of Ca8F16 (pristine) and Ca7ErF16 (doped) were constructed based on the fluorite structure ($a = 5.46$ Å).
   • Electronic Structure: Geometry optimization and ground-state calculations were performed using the GPAW code with the Linear Combination of Atomic Orbitals (LCAO) mode, a DZP basis set, and the PBE exchange-correlation functional.
   • Optical Response: Linear-response TDDFT (LR-TDDFT) within the Casida formalism was used to compute excited-state properties up to 10 eV. Oscillator strengths and transition energies were extracted.
   • Spectrum Generation: Discrete excitations were converted into continuous absorption spectra using Gaussian broadening ($\sigma = 0.1–0.2$ eV) to generate 1,589 energy-resolved data points per system.

2. Feature Engineering:

   • Input variables underwent Box–Cox power transformation to normalize distributions.
   • A linear Support Vector Machine (SVM) was used to rank features by weight magnitude.
   • Three physically interpretable descriptors were selected as the final feature set: the absorption coefficient ($\alpha$), the extinction coefficient ($\kappa$), and the transition energy ($E$).

3. Machine Learning Models:

   • Classical Baseline: A Radial Basis Function (RBF) kernel Support Vector Machine (SVM) was trained on the selected descriptors.
   • Quantum Support Vector Machine (QSVM): A quantum feature map (ZZFeatureMap) was implemented using 3 qubits. The quantum kernel was evaluated on three platforms: an ideal statevector simulator, a noisy QASM-style simulator (depolarizing noise $p=0.05$), and real hardware (IBM Quantum ibm_fez processor).
   • Hybrid Quantum Neural Network (QNN): A variational quantum classifier was constructed using a 3-qubit ZZFeatureMap followed by a depth-4 TwoLocal ansatz (parameterized $R_y$, $R_z$, and CX gates). The quantum output was mapped to class probabilities via a parity function and processed by a classical linear layer. The model was trained end-to-end using the Adam optimizer with early stopping.

Key Results

• Spectral Differences: TDDFT calculations revealed that Er doping induces a systematic red-shift in dominant transitions ($\Delta E \approx -1.5$ to $-0.3$ eV) and a redistribution of oscillator strength, creating distinct optical fingerprints in the visible and near-UV regions compared to the wide-bandgap pristine CaF2.
• Classical Performance: The classical RBF-SVM achieved a test accuracy of 0.983 and an ROC-AUC of 0.999, establishing a high-performance baseline with negligible overfitting.
• QSVM Performance:
  • Ideal Statevector Simulator: Test accuracy of 0.851 (AUC 0.903).
  • Noisy Simulator: Test accuracy of 0.817 (AUC 0.878).
  • Real Hardware (ibm_fez): Test accuracy dropped to 0.733 on a small test slice (15 samples), attributed to decoherence and finite-shot noise affecting the kernel matrix structure.
• Hybrid QNN Performance: The variational QNN achieved a test accuracy of 0.93 and an AUC of 0.96. This performance significantly exceeded the QSVM results and approached the classical baseline, demonstrating that trainable variational parameters can adapt to the feature space geometry more effectively than fixed quantum kernels.

Significance and Claims
The paper positions its work as a systematic benchmarking study rather than a demonstration of definitive quantum advantage. The authors claim that:

1. Robust Feature Space: Dopant-induced optical fingerprints form a physically grounded, robust feature space suitable for benchmarking near-term quantum learning models.
2. QML Evaluation: The study provides a balanced comparison showing that while static quantum kernels (QSVM) capture meaningful class separation, they currently underperform strong classical baselines and are sensitive to hardware noise.
3. Variational Advantage: Hybrid QNNs, through their trainable ansatz, demonstrate superior representational power and flexibility compared to fixed-kernel QSVMs, achieving performance close to classical models even with limited qubits.
4. Real-World Context: The successful execution of the quantum kernel on IBM hardware validates the end-to-end pipeline (embedding, transpilation, execution) but highlights the current limitations of NISQ devices for supervised learning tasks, where noise significantly degrades performance compared to simulations.

The authors conclude that QNN-based architectures offer a promising pathway for adaptable quantum learning in spectral analysis, provided that future developments in error mitigation and hardware fidelity can bridge the gap between simulated potential and real-world deployment.