Held as part of the Bridge Program at AAAI 2026
January 20 to 21, 2026
Singapore EXPO | Singapore
Room: Tourmaline 207 - 209
Scientific knowledge-guided machine learning (KGML) is an emerging field of research where scientific knowledge is deeply integrated in ML frameworks to produce solutions that are scientifically grounded, explainable, and likely to generalize on out-of-distribution samples even with limited training data. By using both scientific knowledge and data as complementary sources of introduction in the design, training, and evaluation of ML models, KGML seeks a distinct departure from black-box data-only methods and holds great potential for accelerating scientific discovery in a number of disciplines.
The goal of this bridge is to nurture the community of researchers working at the intersection of ML and scientific areas and shape the vision of the rapidly growing field of KGML. This bridge builds upon the success of three previous symposiums organized on this topic at the AAAI Fall Symposium Series in 2020, 2021, and 2022. See our book on KGML for a coverage of research topics in this field.
Day 1: January 20
9:00 am to 10.00 am
Introduction to Knowledge-Guided Machine Learning
Speaker: Dr. Anuj Karpatne
10:00 am to 10.30 am
Invited Talk by Dr. Zekun (Danny) Ren
Title: From Physics Priors to Self-Driving Labs: Knowledge-Guided AI for Materials Discovery.
Abstract: Accelerating materials discovery requires integrating physical knowledge with data-driven intelligence across all stages, from atomic design to device optimization. While machine learning excels at pattern recognition, its effectiveness in materials science remains constrained by sparse data, expensive experiments, and the need for physical validity. In this talk, I will present a knowledge-guided framework that unifies physics-grounded representations, generative AI, high-throughput experimentation, and closed-loop optimization into a coherent workflow for materials innovation.
Bio: Dr. Danny Zekun Ren is a junior group lead at Berkeley Education Alliance for Research in Singapore (BEARS), where he leads the development of multi-agent, LLM-driven platforms for electrocatalyst and materials research. He is also Co-founder and head of AI at Xinterra, a startup company using AI to accelerate carbon capture materials commercialization. He has published 50+ international publications, and prior experience as a Principal Investigator at SMART MIT. He earned his Ph.D. in Electrical and Computer Engineering from a joint program between NUS and MIT.
10.30 am to 11.00 am
11:00 am to 11.30 am
Invited Talk by Dr. Amit Sheth
Title: Neurosymbolic approach to intelligent, robust, and trustworthy AI, explained with healthcare and manufacturing applications.
Abstract: In Pedro Domingos' influential 2012 paper, the phrase "Data alone is not enough" emphasized a crucial point. I've long held this belief, as evidenced by our Semantic Search engine, which was commercialized in 2000 with the first patent on the topic. We enhanced machine learning classifiers with a comprehensive WorldModel™, which included knowledge graphs (KGs). This early project highlighted the synergy between data-driven statistical learning and knowledge-supported symbolic AI methods, a key idea driving today’s more scalable approach to NeuroSymbolic AI.
Contemporary GenAI is impressive in many consumer-focused content and creative applications, as well as a growing number of scientific applications. Nevertheless, they fall short for mission-critical applications (e.g., where inaccuracies and the inability to comply with rules, regulations, and guidelines are unacceptable) and for high-value enterprise applications. These applications require stronger, more reliable support for foundational elements of Intelligent, Robust, and Trustworthy (IRT) AI. Some of the 15 foundational elements we have investigated include Reliability, Consistency, Alignment, Causality, Instructability, Reasoning, Planning, Grounding, Interpretability, Explainability, and Safety.
In this talk, we will describe how KG-driven Neurosymbolic AI frameworks deliver enterprise-grade IRT AI, illustrated with examples from healthcare and advanced manufacturing.
Bio: Prof. Amit Sheth is an educator, researcher, and entrepreneur. He is the NCR Chair & Professor of Computer Science & Engg at the University of South Carolina. As the founding director of the university-wide AI Institute of South Carolina (#AIISC), he grew it to nearly 50 researchers and one of the top AI institutes worldwide. He currently leads the ambitious effort to establish the Indian AI Research Organization (IAIRO). He is a fellow of the IEEE, AAAI, AAAS, ACM, and AAIA for his contributions to knowledge-enhanced computing (including neurosymbolic AI), knowledge-based techniques for transforming diverse data into insights and actions, distributed workflows and data semantics, data integration, and related areas. His numerous awards include the IEEE CS Wallace McDowell Award, the IEEE TCSVC Research Innovation Award, and the BITS-Pilani Distinguished Alumnus Award. In 2023 and for a decade prior, he was ranked among the top 100 Computer Scientists in the world (current h-index 124). Three of the AI companies he founded involve licensing his university research outcomes. These include the first Semantic Web company, founded in 1999 and pioneering technology similar to what is found today in Google Semantic Search and Knowledge Graph; a clinical NLP company; and a company (http://cognovilabs.com) at the intersection of emotion and AI. He is particularly proud of the exceptional success of his >55 PhD students and postdocs in academia, industry labs, and entrepreneurships.
11:30 am to 12.00 pm
Abstract: Recent development in deep learning has spurred research advances in time series modeling and analysis. Practical applications of time series raise a series of new challenges in science applications, such as incorporation of physics knowledge, multi-resolution, missing value, uncertainty quantification, and interpretability. In this talk, I will discuss possible paths to foundation models for time series data and future directions in time series.
Bio: Yan Liu is a Professor in the Computer Science Department and the Director of the Machine Learning Center at the University of Southern California. She received her Ph.D. degree from Carnegie Mellon University. Her research interest is machine learning and its applications to climate science, health care, and sustainability. She is a fellow of IEEE and AAAI. She has received several awards, including NSF CAREER Award, Okawa Foundation Research Award, New Voices of Academies of Science, Engineering, and Medicine, Best Paper Award in SIAM Data Mining Conference.
12:00 pm to 12:30 pm
Invited Talk by Dr. Chin Chun Ooi
Title: Physics-Grounded AI: A Discussion of Possibilities from the synergy between Physics, Scientific Computing, and AI.
Abstract: Deep learning has achieved remarkable success in domains abundant with labeled data, yet its application to complex physical systems remains challenged by the high cost of data generation and the need for rigorous adherence to physical laws. Conversely, traditional numerical simulations (e.g., CFD) offer high fidelity but suffer from prohibitive computational costs, limiting their utility in real-time control and iterative design optimization. This talk explores the emerging paradigm of Scientific Machine Learning (SciML), specifically focusing on the synergy possible from the overlapping fields of physics, numerical simulations and deep learning.
Bio: Dr. Ooi Chin Chun is a Research Scientist at the Institute of High Performance Computing (IHPC) and the Centre for Frontier AI Research (CFAR) under A*STAR, Singapore. He received his B.S.E. from Duke University and his Ph.D. in Chemical Engineering from Stanford University (2016). His research lies at the intersection of computational physics and artificial intelligence, specifically within the domain of Scientific Machine Learning (SciML). His work focuses on synergizing the rigorous governing laws of numerical simulations with the computational efficiency of deep learning. He specializes in developing Physics-Informed Machine Learning methodologies to accelerate complex physics simulations, especially in computational fluid dynamics, improve model performance in data-sparse regimes, and enable rapid digital twinning and design optimization for urban sustainability and engineering systems. By integrating physics with neural networks, his work aims to bridge the gap between traditional scientific computing and modern AI, thereby enabling fresh insights across diverse scientific and engineering applications.
12:30 pm to 2:00 pm
2:00 pm to 2.45 pm
Contributed Talks
[2:00pm - 2:08pm] Scientific Knowledge–Guided Machine Learning for Vessel Power Prediction: A Comparative Study.
Presenter: Orfeas Bourchas
[2:08pm - 2:16pm] Exploring LLMs for Scientific Information Extraction using the SciEx Framework.
Presenter: Brian Mayer
[2:16pm - 2:24pm] Task Aware Modulation using Representation Learning for Upsaling of Terrestrial Carbon Fluxes.
Presenter: Arvind Renganathan
[2:24pm - 2:32pm] Knowledge-Guided Masked Autoencoder with Linear Spectral Mixing and Spectral-Angle–Aware Reconstruction.
Presenter: Abdul Matin
[2:32pm - 2:40pm] DECOR: Deep Embedding Clustering with Orientation
Robustness.
Presenter: Fiona Victoria Stanley Jothiraj
[2:40pm - 2:45pm] (Early-Career Track) Knowledge-Guided Priors for Self-Supervised Learning.
Presenter: Navodita Mathur
2:45 pm to 3.30 pm
Poster Session
Poster ID - WS41: Hybrid Lagrangian-Hamiltonian Neural Networks for Physics-Informed Scientific Modeling.
Poster ID - WS42: Beyond Perfect Physics: Analysis of Physics-Informed Neural Networks under Incomplete Scientific Knowledge.
Poster ID - WS43: Predicting Cherry Blossom Peak Bloom in Toronto Through Climate-Aware Tabular Foundation Models.
Poster ID - WS44: Knowledge-Driven Hybrids: The Next Generation of Crop Growth Models.
Poster ID - WS45: Finetune-Informed Pretraining Boosts Downstream Performance.
Poster ID - WS46: TeXpert: A Multi-Level Benchmark for Evaluating LATEX Code Generation by LLMs.
Poster ID - WS47: KG Priors in Self-supervised Learning.
Poster ID - WS48: Bird Classification Through Trait Identification.
3:30 pm to 4.00 pm
4.00 pm to 5.00 pm
Day 2: January 21
9:00 am to 10.30 am
Abstract: Scientific knowledge can appear in a variety of forms, it may be complete or partial and can be integrated into various parts of the deep learning modeling pipeline. In this tutorial, we will highlight and detail some popular techniques to incorporate domain knowledge to improve machine learning model performance.
10:30 am to 11:00 am
11:00 am to 11.30 am
Abstract: Computation has transformed how we study nature. Emerging technologies, such as GPS, high-definition and autonomous imaging, eDNA, bioacoustics, and crowdsourcing, now generate multimodal, multiscale, multisensory data about life on Earth at unprecedented richness. Yet our ability to extract insight from these data still lags behind our ability to collect them. Even the traits of organisms cannot be readily extracted from images. Yet, the analysis of traits, the integrated products of genes and environment, is critical for biologists to understand the function of organisms, predict effects of environmental change or genetic manipulation, and to understand the significance of patterns in the four-billion-year evolutionary history of life.
This talk will showcase how knowledge-guided machine learning and broader AI/ML can turn massive collections of multi-modal data into high resolution information database about living organisms, enabling scientific discovery, conservation, and policy decisions. It will introduce imageomics, a new field of science, and present a vision for AI as a trustworthy partner in scientific discovery.
Bio: Dr. Tanya Berger-Wolf is a Professor of Computer Science Engineering, Electrical and Computer Engineering, and Evolution, Ecology, and Organismal Biology at the Ohio State University, where she is also the Director of the Translational Data Analytics Institute. A pioneer in Artificial Intelligence (AI) for ecology, biodiversity, and conservation, she leads the NSF-funded Imageomics Institute and the US-Canada co-funded AI and Biodiversity Change (ABC) Global Center.
Dr. Berger-Wolf serves as a scientific advisor and board member for many organizations, including the US National Academies Board on Life Sciences, US National Committee for the International Union of Biological Sciences (IUBS), the Global Partnership on AI (GPAI)/OECD, National Ecological Observatory Network (NEON), and The Nature Conservancy. She co-founded the AI for conservation non-profit Wild Me (now part of Conservation X Labs), creator of Wildbook, recognized by UNESCO for advancing the UN Sustainable Development Goals. Her contributions have earned numerous honors, including recognition as the AI 100 Global Thought Leaders by H20.ai and the OSU College of Engineering Lumley Interdisciplinary Research Award. She is an elected Fellow of the Association for the Advancement of Artificial Intelligence (AAAI).
11:30 am to 12:00 pm
Invited Talk by Delia Qu
Title: Unsupervised Combinatorial Probabilistic Reasoning for Scientific Discovery.
Abstract: Over the past decade, artificial intelligence has made remarkable progress through data-centric deep learning in domains such as games, image recognition, speech, and natural language. Nevertheless, scientific tasks pose fundamental challenges due to their unsupervised or weakly supervised nature and the need to adhere to established scientific principles. A central grand challenge is therefore to develop AI systems capable of high-throughput interpretation of scientific data at scales far exceeding the capacity of experts. We introduce the Probabilistic Coin Change Problem (PCCP), a novel variant of the classical Coin Change Problem that captures a broad class of inverse scientific combinatorial tasks, including molecular formula inference from mass spectrometry, a foundational challenge in chemistry with far-reaching implications for biochemistry, medicine, and biology. We present DeepProReasoner, an unsupervised probabilistic reasoning framework for solving PCCP-style problems that substantially outperforms both research and commercial state-of-the-art methods.
Bio: Zhongdi (Delia) Qu is a Ph.D. student in Computer Science at Cornell University. Her current research focuses on developing principled and efficient knowledge-guided machine learning systems and optimization methods, with applications in scientific discovery and computational sustainability. Prior to her doctoral studies, Delia served as a Software Engineer at Google Research in Zurich and Google Search in New York City. During her time in industry, she developed machine learning models for zero-shot semantic parsing and end-to-end spoken language understanding. She holds a Master's degree in Computer Science from École Normale Supérieure Paris-Saclay and a B.S. in Computer Science and Mathematics from Vanderbilt University.
12:30 pm to 2:00 pm
2:00 pm to 2:30 pm
Invited Talk by Dr. Zhenong Jin
Title: Knowledge-Guided Data Assimilation with Agroecosystem Digital Twins
Abstract: The development of agricultural digital twins offers a transformative path toward sustainable food production systems. This presentation explores how Knowledge-Guided Machine Learning (KGML) facilitates this vision by integrating physical models and expert knowledge into deep learning frameworks for vegetation monitoring. We highlight two key advances in estimating canopy properties. The first case demonstrates a pre-training and fine-tuning Bi-LSTM framework that improves the generation of high spatiotemporal resolution Leaf Area Index (LAI) products. The second case introduces an innovative method combining gap-fraction theory with Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). This approach estimates LAI from 2D images by creating high-fidelity 3D digital twins, bridging the gap between point-level measurements and large-scale monitoring. The estimated canopy properties further improved a 3D biophysical model of crop photosynthesis, a key physical process regulating the carbon fluxes from croplands. Furthermore, we introduce a prototype of ecological digital twins that can be used to extract and encode domain knowledge to guide these complex data assimilation processes. Collectively, these hybrid models closes the gap between vegetation’s morphology and function modeling, laying the groundwork for next-generation tools in precision agriculture and sustainable agroecosystem management.
Bio: Zhenong Jin is a tenured Associate Professor at the College of Urban and Environmental Sciences, Peking University. His experience also involves research, teaching, and product development at Stanford University, Atlas AI, and University of Minnesota. His research mainly focuses on integrating ecological knowledge, process-based modelling, remote sensing, and machine learning to provide scientific and technical foundation for monitoring and managing the food production systems. His papers have been extensively published in leading journals such as Science, Nature Climate Change, Nature Food, Nature Geoscience, and Remote Sensing of Environment. He is a recipient of the NSF CAREER Award in the United States.
2:30 pm to 3:00 pm
Abstract: Physics-informed machine learning (PIML) has emerged as a powerful paradigm for integrating first-principles knowledge with data-driven models, offering a promising pathway for modeling complex chemical processes under limited, noisy, or heterogeneous data. In this talk, we present recent advances in physics-informed neural networks (PINNs) for constructing dynamic process models governed by ordinary differential equations (ODEs). We address both forward and inverse modeling problems across a range of chemical engineering applications, including scenarios with partially known process models, and discuss strategies for enforcing physical constraints during training.
A key practical challenge in PINN training lies in balancing multiple competing loss terms, including data fidelity, physics residuals, and boundary or initial conditions. I will show how this challenge can be systematically addressed using multi-objective optimization (MOO), leading to improved training stability, better convergence behavior, and more interpretable trade-offs between data consistency and physical law enforcement. The methodology discussed in this talk will be demonstrated through several simulation and experimental case studies, including nonlinear chemical reaction systems and a photochemical reactor.
Bio: Zhe Wu is an Assistant Professor in the Department of Chemical and Biomolecular Engineering at the National University of Singapore (NUS). He earned his B.S. in Control Science and Engineering from Zhejiang University (2016) and his Ph.D. in Chemical Engineering from UCLA (2020), followed by a postdoctoral position in UCLA’s Computer Science department. His research focuses on process systems engineering, including process modeling, optimization, control, and machine learning applications in chemical engineering. He has published over 50 peer-reviewed articles and secured more than SGD 2 million in research funding from Singapore funding agencies: ASTAR, NRF, MOE, and chemical and pharmaceutical companies. Dr. Wu serves on the editorial board of Digital Chemical Engineering and is on the Early Career Advisory Board of Control Engineering Practice. He has received the Outstanding Early Career Award from NUS (2024), the Outstanding Young Faculty Award from AIChE Singapore Local Section (2024), and was named to the Stanford/Elsevier Top 2% Scientists List (2024).
3:00 pm to 3:30 pm
Abstract: Modern industrial systems generate vast streams of sensor data that capture the behavior and health of complex physical assets. While machine learning has become a powerful tool for analyzing such data, purely data-driven approaches often struggle with limited labeled examples, lack of interpretability, and poor generalization to rare or previously unseen failure conditions. This talk explores how scientific knowledge-guided machine learning can address these challenges by embedding established engineering knowledge, such as reliability principles and failure analysis, directly into learning frameworks for industrial asset health monitoring. By treating domain knowledge and data as complementary sources of supervision, knowledge-guided approaches enable models to reason about assets, sensors, and failure mechanisms in a way that is explainable, data-efficient, and robust to distribution shifts. Through examples from industrial monitoring, the talk highlights a broader vision of KGML as a bridge between machine learning and engineering science.
Bio: Shuxin Lin is a Research Scientist at IBM Research in Yorktown Heights, New York. She received her M.S. degree from Carnegie Mellon University in 2018. Her research focuses on building reliable and scalable AI systems, with recent interests in datasets and benchmarks, post-training of LLMs, and LLM agentic systems. Her work spans both foundational methods and real-world applications in AI automation in the domain of Industry 4.0.
3:30 pm to 4:00 pm
4.00 pm to 5.00 pm
5.00 pm to 5.15 pm
Nikhil Muralidhar
Stevens Institute of Technology nmurali1@stevens.edu
Taniya Kapoor
Wageningen University & Research
taniya.kapoor@wur.nl
Kai-Hendrik Cohrs
Universitat de València
kai.cohrs@uv.es
Ramakrishnan Kannan
Oak Ridge National Laboratory
kannanr@ornl.gov
Anuj Karpatne
Virginia Tech
karpatne@vt.edu
Xiaowei Jia
University of Pittsburgh
xiaowei@pitt.edu
Vipin Kumar
University of Minnesota
kumar001@umn.edu
