Min Xian

Metaphor requires a language model to resolve a token whose contextual meaning diverges from its basic literal sense. Understanding how transformer models organize this reinterpretation across depth remains an open problem in mechanistic interpretability. We introduce conditional scale entropy (CSE), a wavelet-derived measure of how broadly transformer computation engages across frequency scales at each layer position. Two theorems establish that CSE is invariant to update magnitude, isolating the structural pattern of updates from their intensity. Using CSE, we find that metaphorical tokens produce significantly higher spectral breadth than literal tokens at contiguous layer positions on every decoder-only architecture tested, from 124M to 20B parameters (GPT-2 family, LLaMA-2 7B, GPT-oss 20B). The effect survives cluster-based permutation correction, recurs in the early-to-mid relative depth range across models, and converges with an independent analysis of 200 naturalistic VUA pairs. Specificity controls further show that the effect is not explained by semantic complexity or by matched propositional content. These results identify multi-scale coordination as a consistent signature of metaphorical language processing in the decoder-only architectures examined, and establish CSE as a principled tool for characterizing cross-depth structure in transformers.

Tristructural isotropic (TRISO)-coated particle fuels undergo dimensional changes and chemical reactions during high-temperature neutron irradiation. Post-irradiation materialography helps understand processes that impact fuel performance, such as coating integrity and fission product retention. Conventionally, experts manually evaluate features in thousands of cross sections of sub-mm-sized samples, which is tedious and subjective. In this work, we propose UA-Net, a deep learning framework that segments five characteristic regions of TRISO fuel micrographs and generates an uncertainty map for predictions. The model uses a multi-stage pretraining strategy, starting with general image representations learned from ImageNet, followed by fine-tuning on TRISO micrographs from various irradiation experiments and AGR-5/6/7 particle cross sections. A meta-model for uncertainty prediction is integrated to identify small defects in TRISO images. UA-Net was evaluated on a test set of 102 images, achieving mean Intersection over Union (mIoU) and mean Precision (mP) of 95.5% and 97.3%, respectively. The meta-model achieved a specificity of 91.8% and sensitivity of 93.5%, demonstrating strong performance in detecting misclassifications. The model was also applied to new TRISO images for qualitative evaluation, showing high accuracy in extracting layer regions.

Failure event narratives contain detailed and valuable information describing how failures initiate and propagate. Event causality analysis can help improve the understanding of failure physics and facilitate the use of non-failure data (e.g., near-misses and degradations) to complement the limited data pool of failures, which is common in high-reliability industries such as the nuclear power industry. Automatically extracting event causality from text data, however, is challenging given complex and diverse language structures and causal patterns, and the lack of access to large, annotated datasets for use as training data. Existing automated mining approaches are mainly knowledge-based and extract causality using a set of predefined keywords and rules, which have difficulty achieving good performance. In this paper, we propose a novel large language model (LLM)-based approach for automated causality extraction. It leveraged the strong capability of LLM to understand intricate language patterns in long-range contexts and accurately extract cause-and-effect pairs from texts. The proposed approach has a twofold framework: causality detection and causality extraction. The causality detection step trained a deep learning model to identify texts with causality. The causality extraction step developed a T5-CE LLM to identify and extract cause-and-effect pairs in each text sample. A large, annotated dataset of the U.S. nuclear power plant failure event reports was used to train and evaluate the models. The model evaluation was performed using three performance metrics, including precision, recall, and F1 score. The proposed approach can effectively detect implicit and embedded causalities across multiple sentences.

During irradiation, phenomena such as kernel swelling and buffer densification impact the performance of tristructural isotropic (TRISO) particle fuel. Post-irradiation microscopy is often used to characterize the morphological changes induced by irradiation. However, each fuel compact generally contains thousands of TRISO particles. Manually collecting data to obtain quantitative characterizations of these phenomena is cumbersome and subjective. To address the challenges, we developed a convolutional neural network (CNN), namely RU-Net, to accelerate the characterization of TRISO fuel cross sections. We built a large dataset of irradiated TRISO particles, comprising 2171 microscopic images of cross-sectioned particles and their corresponding annotations. The proposed RU-Net has a two-encoder design that extracts and fuses image context at different scales and accurately segments TRISO layers of varying sizes. Extensive experiments have been conducted on the proposed large dataset to evaluate the performance of the RU-Net and other state-of-the-art CNNs. The results demonstrated that the proposed RU-Net achieved the best overall performance on the test set. Using the results of RU-Net segmentation, we can expedite analysis of TRISO particle cross sections, significantly reducing manual labor and improving the objectivity of the results.

Identifying and classifying shutdown initiating events (SDIEs) is critical for developing shutdown probabilistic risk assessment for nuclear power plants. Existing computational approaches cannot achieve satisfactory performance due to the challenges of unavailable large, labeled datasets, imbalanced event types, and label noise. To address these challenges, we propose a hybrid pipeline that integrates a knowledge-informed machine learning model to prescreen non-SDIEs and a large language model (LLM) to classify SDIEs into four types. In the prescreening stage, we proposed a set of 44 SDIE text patterns that consist of the most salient keywords and phrases from six SDIE types. Text vectorization based on the SDIE patterns generates feature vectors that are highly separable by using a simple binary classifier. The second stage builds Bidirectional Encoder Representations from Transformers (BERT)-based LLM, which learns generic English language representations from self-supervised pretraining on a large dataset and adapts to SDIE classification by fine-tuning it on an SDIE dataset. The proposed approaches are evaluated on a dataset with 10,928 events using precision, recall ratio, F 1 score, and average accuracy. The results demonstrate that the prescreening stage can exclude more than 97% non-SDIEs, and the LLM achieves an average accuracy of 95.1% for SDIE classification.

During irradiation, phenomena such as kernel swelling and buffer densification may impact the performance of tristructural isotropic (TRISO) particle fuel. Post-irradiation microscopy is often used to identify these irradiation-induced morphologic changes. However, each fuel compact generally contains thousands of TRISO particles. Manually performing the work to get statistical information on these phenomena is cumbersome and subjective. To reduce the subjectivity inherent in that process and to accelerate data analysis, we used convolutional neural networks (CNNs) to automatically segment cross-sectional images of microscopic TRISO layers. CNNs are a class of machine-learning algorithms specifically designed for processing structured grid data. They have gained popularity in recent years due to their remarkable performance in various computer vision tasks, including image classification, object detection, and image segmentation. In this research, we generated a large irradiated TRISO layer dataset with more than 2,000 microscopic images of cross-sectional TRISO particles and the corresponding annotated images. Based on these annotated images, we used different CNNs to automatically segment different TRISO layers. These CNNs include RU-Net (developed in this study), as well as three existing architectures: U-Net, Residual Network (ResNet), and Attention U-Net. The preliminary results show that the model based on RU-Net performs best in terms of Intersection over Union (IoU). Using CNN models, we can expedite the analysis of TRISO particle cross sections, significantly reducing the manual labor involved and improving the objectivity of the segmentation results.

Artificial Intelligence (AI)'s capacity to analyze dermoscopic images promises a ground-breaking leap in skin cancer diagnostics, offering exceptional accuracy and a effortlessly non-invasive image acquisition process. However, this immense potential, which has ignited widespread research enthusiasm, is critically undermined due to the lack of transparency and interpretability. The automated generation of articulate and comprehensive diagnostic reports will bridge this critical gap by not only illuminate the AI's diagnostic rational but also substantially reduce the demanding workload of the medical professionals. This study presents a multimodal vision-language model (VLM) trained using a two-stage knowledge distillation (KD) framework to generate structured medical reports from dermoscopic images, with descriptive features based on the 7-point melanoma checklist. The reports are organized into clinically relevant sections-Findings, Impression, and Differential Diagnosis-aligned with dermatological standards. Experimental evaluation demonstrates the system's ability to produce accurate and interpretable reports. Human feedback from a medical professional, assessing clinical relevance, completeness, and interpretability, supports the utility of the generated reports, while computational metrics validate their accuracy and alignment with reference pseudo-reports, achieving a SacreBLEU score of 55.59, a ROUGE-1 score of 0.5438, a ROUGE-L score of 0.3828, and a BERTScore F1 of 0.9025. These findings underscore the model's ability to generalize effectively to unseen data, enabled by its multimodal design, clinical alignment, and explainability.

Producing large images using small diffusion models is gaining increasing popularity, as the cost of training large models could be prohibitive. A common approach involves jointly generating a series of overlapped image patches and obtaining large images by merging adjacent patches. However, results from existing methods often exhibit obvious artifacts, e.g., seams and inconsistent objects and styles. To address the issues, we proposed Guided Fusion (GF), which mitigates the negative impact from distant image regions by applying a weighted average to the overlapping regions. Moreover, we proposed Variance-Corrected Fusion (VCF), which corrects data variance at post-averaging, generating more accurate fusion for the Denoising Diffusion Probabilistic Model. Furthermore, we proposed a one-shot Style Alignment (SA), which generates a coherent style for large images by adjusting the initial input noise without adding extra computational burden. Extensive experiments demonstrated that the proposed fusion methods improved the quality of the generated image significantly. As a plug-and-play module, the proposed method can be widely applied to enhance other fusion-based methods for large image generation.

Recent advancements in machine learning-based methods have demonstrated great
potential for improved property prediction in material science. However,
reliable estimation of the confidence intervals for the predicted values
remains a challenge, due to the inherent complexities in material modeling.
This study introduces a novel approach for uncertainty quantification in
fatigue life prediction of metal materials based on integrating knowledge from
physics-based fatigue life models and machine learning models. The proposed
approach employs physics-based input features estimated using the Basquin
fatigue model to augment the experimentally collected data of fatigue life.
Furthermore, a physics-informed loss function that enforces boundary
constraints for the estimated fatigue life of considered materials is
introduced for the neural network models. Experimental validation on datasets
comprising collected data from fatigue life tests for Titanium alloys and
Carbon steel alloys demonstrates the effectiveness of the proposed approach.
The synergy between physics-based models and data-driven models enhances the
consistency in predicted values and improves uncertainty interval estimates.

The generalization performance of deep neural networks (DNNs) is a critical
factor in achieving robust model behavior on unseen data. Recent studies have
highlighted the importance of sharpness-based measures in promoting
generalization by encouraging convergence to flatter minima. Among these
approaches, Sharpness-Aware Minimization (SAM) has emerged as an effective
optimization technique for reducing the sharpness of the loss landscape,
thereby improving generalization. However, SAM's computational overhead and
sensitivity to noisy gradients limit its scalability and efficiency. To address
these challenges, we propose Gradient-Centralized Sharpness-Aware Minimization
(GCSAM), which incorporates Gradient Centralization (GC) to stabilize gradients
and accelerate convergence. GCSAM normalizes gradients before the ascent step,
reducing noise and variance, and improving stability during training. Our
evaluations indicate that GCSAM consistently outperforms SAM and the Adam
optimizer in terms of generalization and computational efficiency. These
findings demonstrate GCSAM's effectiveness across diverse domains, including
general and medical imaging tasks. Our code is available at https://github.com/mhassann22/GCSAM