Machine Learning Algorithms for Quantum System Modeling

OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems Accurately modeling chemical systems across diverse charges, spin states, and environments remains a central challenge in molecular machine learning. No existing machine learning–based methods can simultaneously handle molecules with varying charges, spins, and environments. A few recently developed approaches address one or two of these factors individually by designing task-specific architectures, but this limits their applicability to broader chemical scenarios. OrbitAll is the first deep learning-based method that can simultaneously incorporate spin, charge, and environmental information using consistent and physically grounded quantum mechanical features. It has superior accuracy, generalization, and data efficiency on diverse chemical systems. We introduce a unified quantum mechanical representation that naturally incorporates spin, charge, and environmental effects within a single, physics-informed framework. Specifically, OrbitAll utilizes spin-polarized orbital features from the underlying quantum mechanical method, and combines it with graph neural networks satisfying SE(3)-equivariance. This enables our model, OrbitAll, to achieve accurate, robust, and data-efficient predictions across a wide range of chemical systems–including charged and open-shell species, as well as solvated molecules–without the need for domain-specific tuning. OrbitAll achieves chemical accuracy using 10 times fewer training data than competing AI models, with a speedup of more than thousand times compared to density functional theory. It can extrapolate to molecules more than 10times larger than those in training data. This universality distinguishes our approach from current deep learning models.

I’d like to draw your attention to a new paper on arXiv, “Shallow-circuit Supervised Learning on a Quantum Processor”, from IBM and Qognitive that develops a Hamiltonian-based framework for quantum machine learning. Instead of fixed amplitude or angle encodings used in many prior approaches, our method learns a local Hamiltonian embedding for classical data. https://lnkd.in/ejcxYstW We are very interested in new approaches to QML as we deal with recurring bottlenecks like expensive classical data loading and difficult training dynamics in parameterized circuit models. Here, both the feature operators and the label operator are learned during training, with predictions obtained from measurements on an approximate ground state. This aims to avoid those bottlenecks. A key enabler is Sample-based Krylov Quantum Diagonalization (SKQD), which approximates low-energy states by sampling from time-evolved Krylov states and then diagonalizing the Hamiltonian in the sampled subspace. SKQD was recently employed to estimate low-energy properties of impurity models (https://lnkd.in/epwCrG5R). In our setting, restricting to 2-local Hamiltonian embeddings keeps the required time-evolution circuits relatively shallow, which helps make the approach practical on current quantum processors. The team demonstrates end-to-end training on IBM Heron processor up to 50 qubits, with non-vanishing gradients and strong proof-of-concept performance on a binary classification task. There are many exciting next steps here, including testing on broader datasets, using more expressive operator ansatz, and performing systematic comparisons to strong classical baselines to pinpoint when Hamiltonian-based encodings offer the right inductive bias. I encourage the community to try out this approach and explore where it can be extended in meaningful ways.

Interesting new study: "EnQode: Fast Amplitude Embedding for Quantum Machine Learning Using Classical Data." The authors introduce a novel framework to address the limitations of traditional amplitude embedding (AE) [GitHub repo included]. Traditional AE methods often involve deep, variable-length circuits, which can lead to high output error due to extensive gate usage and inconsistent error rates across different data samples. This variability in circuit depth and gate composition results in unequal noise exposure, obscuring the true performance of quantum algorithms. To overcome these challenges, the researchers developed EnQode, a fast AE technique based on symbolic representation. Instead of aiming for exact amplitude representation for each sample, EnQode employs a cluster-based approach to achieve approximate AE with high fidelity. Here are some of the key aspects of EnQode: * Clustering: EnQode begins by using the k-means clustering algorithm to group similar data samples. For each cluster, a mean state is calculated to represent the central characteristics of the data distribution within that cluster. * Hardware-optimized ansatz: For each cluster's mean state, a low-depth, machine-optimized ansatz is trained, tailored to the specific quantum hardware being used (e.g., IBM quantum devices). * Transfer Learning for fast embedding: Once the cluster models are trained offline, transfer learning is used for rapid amplitude embedding of new data samples. An incoming sample is assigned to the nearest cluster, and its embedding circuit is initialized with the optimized parameters of that cluster's mean state. These parameters can then be fine-tuned, significantly accelerating the embedding process without retraining from scratch. * Reduced circuit complexity: EnQode achieved an average reduction of over 28× in circuit depth, over 11× in single-qubit gate count, and over 12× in two-qubit gate count, with zero variability across samples due to its fixed ansatz design. * Higher state fidelity in noisy environments: In noisy IBM quantum hardware simulations, EnQode showed a state fidelity improvement of over 14× compared to the baseline, highlighting its robustness to hardware noise. While the baseline achieved 100% fidelity in ideal simulations (as it performs exact embedding), EnQode maintained an average of 89% fidelity when transpiled to real hardware in ideal simulations, which is considered a good approximation given the significant reduction in circuit complexity. Here the article: https://lnkd.in/dQMbNN7b And here the GitHub repo: https://lnkd.in/dbm7q3eJ #qml #datascience #machinelearning #quantum #nisq #quantumcomputing

"Discovering chemical reaction pathways using quantum mechanics is impractical for many systems of practical interest because of unfavorable scaling and computational cost. While machine learning interatomic potentials (MLIPs) trained on quantum mechanical data offer a promising alternative, they face challenges for reactive systems due to the need for extensive sampling of the potential energy surface in regions that are far from equilibrium geometries. Unfortunately, traditional MLIP training protocols are not designed for comprehensive reaction exploration. We present a reactive active learning (RAL) framework that is designed to efficiently train MLIPs to achieve near-quantum mechanical accuracy for reactive systems for situations where one may not have prior knowledge of the possible transition states, reaction pathways, or even the potential products. Our method combines automated reaction exploration, uncertainty-driven active learning, and transition state sampling to build accurate potentials." "We demonstrated that the RAL framework can efficiently train MLIPs to account for chemical reactions with near-[density functional theory] DFT accuracy. RAL differs from other training methods in that it uses methods to automatically generate reaction products based only on specifying reactants and then automatically produces commands to use partially trained MLIPs to explore reaction pathways through reaction-explicit active learning. Reaction pathways are explicitly sampled through transition state finding methods, such as [single-ended growing string method] SE-GSM, using the MLIP that is being trained to perform reaction exploration AL. We have shown that identifying just a few possible reaction products can lead to hundreds of individual reaction pathways for use in RAL. Reaction intermediates and products are automatically discovered, and MLIPs are trained with MD-AL to learn these new chemical species during the RAL process. Thus, the user does not need to provide predefined pathways or even a list of potential products." https://lnkd.in/eiHdaSaD