Gleb Vitalevich Solovev, Alina Borisovna Zhidkovskaya, Anastasia Orlova, Nina Gubina, Anastasia Vepreva, Rodion Golovinskii, Ilya Tonkii, Ivan Dubrovsky, Ivan Gurev, Dmitry Gilemkhanov, Denis Chistiakov, Timur A. Aliev, Ivan Poddiakov, Galina Zubkova, Ekaterina V. Skorb, Vladimir Vinogradov, Alexander Boukhanovsky, Nikolay Nikitin, Andrei Dmitrenko, Anna Kalyuzhnaya, Andrey Savchenko. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025.
Full-penetration laser welding (FPLW) is increasingly adopted in manufacturing pipelines, yet its industrial scalability is constrained by in-process defect formation, particularly incomplete penetration. To address this, we propose a sensor-driven framework for non-destructive monitoring and automated defect detection that uses infrared (IR) thermography as the primary in situ sensing modality and applies deep learning to the acquired thermal signals. High-speed IR camera recordings were processed to track spatter and the weld zone, yielding a time series of physically interpretable spatiotemporal features (mean spatter area, mean spatter temperature, number of spatters, and mean welding zone temperature). Defect recognition is formulated as a multi-label classification problem targeting incomplete penetration, sagging, shrinkage groove, and linear misalignment, and multiple temporal models were evaluated on the same sensor-derived feature sequences. Experimental validation on 09G2S pipeline steel demonstrates that the proposed time series pipeline based on a hybrid CNN–transformer achieves a mean Average Precision (mAP) of 0.85 while preserving near-real-time inference on a CPU. The results indicate that IR thermography-based spatter dynamics provide actionable sensing signatures for automated defect prediction and can serve as a foundation for closed-loop quality control in industrial laser pipeline welding.
Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly large language models (LLMs), have enabled virtual screening methods that reduce costs and improve efficiency. However, the growing complexity of these tools has limited their accessibility to wet-lab researchers. Multi-agent systems offer a promising solution by combining the interpretability of LLMs with the precision of specialized models and tools. In this work, we present MADD, a multi-agent system that builds and executes customized hit identification pipelines from natural language queries. MADD employs four coordinated agents to handle key subtasks in de novo compound generation and screening. We evaluate MADD across seven drug discovery cases and demonstrate its superior performance compared to existing LLM-based solutions. Using MADD, we pioneer the application of AI-first drug design to five biological targets and release the identified hit molecules. Finally, we introduce a new benchmark of query-molecule pairs and docking scores for over three million compounds to contribute to the agentic future of drug design.
Modern industrial tasks, such as quality control in laser welding and the localization of geometric features in industrial processes, require the application of innovative machine learning approaches. The scarcity of annotated data and the complexity of geometric annotation are critical barriers to the development of automated inspection systems. The scientific novelty of the proposed approach lies in the comprehensive use of hybrid methods, combining evolutionary optimization, diffusion models, and convolutional neural networks to effectively address practical engineering tasks with limited data resources. The proposed framework consists of two integrated components. The first component implements a hybrid algorithm for synthetic data generation, merging evolutionary optimization for generating diverse geometric variants with diffusion models for synthesizing photorealistic images. The second component involves a specialized deep learning architecture optimized for the precise localization and classification of geometric features in industrial settings. Training is performed using a combined loss function that integrates regression and classification criteria. In the case of laser welding quality control, the synthetic dataset was expanded from 120 original images to 4,537 realistic samples. This augmentation improved weld seam segmentation accuracy, reducing the box loss metric from 2.4 to 0.75. For the task of localizing weld seam coordinates, a prediction error of 31.8 pixels along the Y-axis and 3.3 pixels along the X-axis was achieved at the original resolution of 1,024 × 2,448 pixels. Experimental comparisons showed that convolutional architectures outperformed transformer-based models with a comparable number of parameters, and that regression from a single frame yielded higher accuracy than using a sequence of frames. The proposed methods demonstrate significant superiority over classical data augmentation techniques (e.g., mixup, cutmix) and pure diffusionbased synthesis approaches which require intensive dataset preparation. The integration of evolutionary optimization ensures controlled diversity in geometric variants, while diffusion models guarantee the photorealism of synthesized samples. This hybrid approach holds broad potential for application in other industrial sectors with limited availability of annotated data, owing to its capability to construct a complete pipeline for synthesizing hard-to-obtain industrial data and subsequently using it to train applied Artificial Intelligence methods for solving targeted industrial problems.