Selected work

Representative Papers

Autonomous battery research: Principles of heuristic operando experimentation

Dec 29, 2025

Traditional in situ battery characterization methods struggle to reliably capture stochastic, transient failure events such as dendrite initiation. This work proposes a heuristic in situ experimental framework that integrates physics-informed digital twins with AI agents to actively guide multimodal beamline instrumentation toward mechanistically critical precursors. Departing from conventional uncertainty-driven active learning, the approach innovatively employs entropy-based metrics to quantify scientific information gain, thereby enhancing experimental efficiency and data value while adhering to FAIR data principles. The method effectively mitigates beam-induced damage and data redundancy, successfully capturing transient precursor phenomena overlooked by conventional techniques, and establishes a new paradigm for building trustworthy autonomous battery laboratories.

1 citationsRead paper

Quantum Reservoir Computing for Short-Term Power Load Forecasting in Resource-Constrained Energy Systems

Jun 10, 2026

This work addresses the challenge of achieving high-accuracy short-term electric load forecasting on edge devices under constraints of limited memory, restricted measurement budgets, and hardware-induced noise. The authors propose a hybrid architecture combining a fixed quantum reservoir with a classical Elastic Net readout layer and, for the first time, integrate post-training low-bitwidth fixed-point quantization (2–8 bits) into the readout stage of quantum reservoir computing. Experimental results on the Tetouan and Spain datasets demonstrate that 6-bit quantization preserves full-precision performance while reducing readout memory usage by 81.2%. Although lower bitwidths induce slight performance degradation—exhibiting dataset-dependent behavior—the model maintains strong transferability under IBM’s quantum noise model without requiring retraining.

0 citationsRead paper

Does Continued Pretraining on a Learner Corpus Improve Automated Essay Scoring on English Proficiency Tests? Evidence from EFCAMDAT

May 25, 2026

This study addresses the limitations of general-purpose pretrained language models in effectively capturing the linguistic characteristics of second-language learners’ writing, which constrains the performance of automated essay scoring (AES). To mitigate this issue, the authors propose a domain-adaptive continued pretraining (DAPT) strategy that leverages CEFR-level filtering to align pretraining data with target proficiency levels. Specifically, three Transformer-based encoders are further pretrained on the EFCAMDAT corpus and evaluated on the FCE and IELTS datasets for both scoring accuracy and cross-dataset transferability. Experimental results demonstrate significant improvements in AES performance when the CEFR levels of the pretraining data match those of the downstream task—such as B1–B2 for FCE—whereas cross-dataset transfer yields inconsistent gains, underscoring the critical importance of proficiency-level alignment in model adaptation.

0 citationsRead paper

Late Breaking Results: Hardware-Efficient Quantum Reservoir Computing via Quantized Readout

Apr 07, 2026

This work addresses the challenge of achieving high-accuracy short-term electric load forecasting on resource-constrained edge devices by proposing a fixed-structure quantum reservoir computing approach. The method integrates Chebyshev feature encoding, a brickwall-type entangling circuit, and Pauli measurements to circumvent the need for quantum backpropagation. Innovatively, post-training fixed-point quantization is introduced in the readout layer, and a genetic algorithm is employed to optimize the overall architecture. Evaluated on the Tetouan city dataset, 8-bit and 6-bit quantization schemes reduce memory usage by 75% and 81%, respectively, while limiting prediction accuracy degradation to within 1% of the FP32 baseline, thereby significantly enhancing feasibility for edge deployment.

0 citationsRead paper

Benchmarking Autonomy in Scientific Experiments: A Hierarchical Taxonomy for Autonomous Large-Scale Facilities

Jan 11, 2026

This work addresses the absence of standardized evaluation criteria for autonomous scientific experimentation in large-scale user facilities, where existing taxonomies rely on an owner-operator model that is ill-suited to such environments. The authors propose the BASE scale—a six-level (0–5) autonomy classification framework tailored for these facilities—that introduces “reasoning barrier” (Level 3) as a critical threshold, marking the transition from scalar feedback-based decisions to those enabled by semantic digital twins and time-gated mechanisms. Integrating hierarchical architectures, real-time inference, and time-synchronized technologies, the framework supports zero-shot deployment of intelligent agents. It provides facility managers, funding agencies, and scientists with a standardized metric to assess risk, delineate responsibility, and quantify the degree of intelligence embedded in experimental workflows.

0 citationsRead paper

Recent publications

Latest Papers

Quantum Reservoir Computing for Short-Term Power Load Forecasting in Resource-Constrained Energy Systems

Jun 10, 2026

This work addresses the challenge of achieving high-accuracy short-term electric load forecasting on edge devices under constraints of limited memory, restricted measurement budgets, and hardware-induced noise. The authors propose a hybrid architecture combining a fixed quantum reservoir with a classical Elastic Net readout layer and, for the first time, integrate post-training low-bitwidth fixed-point quantization (2–8 bits) into the readout stage of quantum reservoir computing. Experimental results on the Tetouan and Spain datasets demonstrate that 6-bit quantization preserves full-precision performance while reducing readout memory usage by 81.2%. Although lower bitwidths induce slight performance degradation—exhibiting dataset-dependent behavior—the model maintains strong transferability under IBM’s quantum noise model without requiring retraining.

0 citationsRead paper

Does Continued Pretraining on a Learner Corpus Improve Automated Essay Scoring on English Proficiency Tests? Evidence from EFCAMDAT

May 25, 2026

This study addresses the limitations of general-purpose pretrained language models in effectively capturing the linguistic characteristics of second-language learners’ writing, which constrains the performance of automated essay scoring (AES). To mitigate this issue, the authors propose a domain-adaptive continued pretraining (DAPT) strategy that leverages CEFR-level filtering to align pretraining data with target proficiency levels. Specifically, three Transformer-based encoders are further pretrained on the EFCAMDAT corpus and evaluated on the FCE and IELTS datasets for both scoring accuracy and cross-dataset transferability. Experimental results demonstrate significant improvements in AES performance when the CEFR levels of the pretraining data match those of the downstream task—such as B1–B2 for FCE—whereas cross-dataset transfer yields inconsistent gains, underscoring the critical importance of proficiency-level alignment in model adaptation.

0 citationsRead paper

Late Breaking Results: Hardware-Efficient Quantum Reservoir Computing via Quantized Readout

Apr 07, 2026

This work addresses the challenge of achieving high-accuracy short-term electric load forecasting on resource-constrained edge devices by proposing a fixed-structure quantum reservoir computing approach. The method integrates Chebyshev feature encoding, a brickwall-type entangling circuit, and Pauli measurements to circumvent the need for quantum backpropagation. Innovatively, post-training fixed-point quantization is introduced in the readout layer, and a genetic algorithm is employed to optimize the overall architecture. Evaluated on the Tetouan city dataset, 8-bit and 6-bit quantization schemes reduce memory usage by 75% and 81%, respectively, while limiting prediction accuracy degradation to within 1% of the FP32 baseline, thereby significantly enhancing feasibility for edge deployment.

0 citationsRead paper

Benchmarking Autonomy in Scientific Experiments: A Hierarchical Taxonomy for Autonomous Large-Scale Facilities

Jan 11, 2026

This work addresses the absence of standardized evaluation criteria for autonomous scientific experimentation in large-scale user facilities, where existing taxonomies rely on an owner-operator model that is ill-suited to such environments. The authors propose the BASE scale—a six-level (0–5) autonomy classification framework tailored for these facilities—that introduces “reasoning barrier” (Level 3) as a critical threshold, marking the transition from scalar feedback-based decisions to those enabled by semantic digital twins and time-gated mechanisms. Integrating hierarchical architectures, real-time inference, and time-synchronized technologies, the framework supports zero-shot deployment of intelligent agents. It provides facility managers, funding agencies, and scientists with a standardized metric to assess risk, delineate responsibility, and quantify the degree of intelligence embedded in experimental workflows.

0 citationsRead paper

Autonomous battery research: Principles of heuristic operando experimentation

Dec 29, 2025

Traditional in situ battery characterization methods struggle to reliably capture stochastic, transient failure events such as dendrite initiation. This work proposes a heuristic in situ experimental framework that integrates physics-informed digital twins with AI agents to actively guide multimodal beamline instrumentation toward mechanistically critical precursors. Departing from conventional uncertainty-driven active learning, the approach innovatively employs entropy-based metrics to quantify scientific information gain, thereby enhancing experimental efficiency and data value while adhering to FAIR data principles. The method effectively mitigates beam-induced damage and data redundancy, successfully capturing transient precursor phenomena overlooked by conventional techniques, and establishes a new paradigm for building trustworthy autonomous battery laboratories.

1 citationsRead paper