Our online joint seminars are at the intersection of AI and fundamental physics open to all participants. We are open to all interested participants across East Asia and beyond. For zoom links, please check out our Slack channel or contact organizers.

Upcoming Seminars

July 23 (Thu)

KalmanAI: Statistical Intelligence for AI-Assisted Scientific Reconstruction

#AI4KMI seminar

  • Speaker: Masako Iwasaki (OMU)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • Statistical inference has long provided the foundation for scientific reconstruction, enabling principled estimation and uncertainty propagation across a wide range of applications. With the rapid adoption of artificial intelligence in scientific computing, an important question arises: how can modern AI be integrated within statistically consistent inference rather than replacing it? This seminar introduces the motivation and design principles behind KalmanAI, an ongoing effort toward a modular Engine-Model-Filter architecture for sequential statistical inference. After revisiting the Bayesian foundations underlying Kalman filtering, the discussion will illustrate how these principles naturally extend across diverse estimation problems, including spacecraft navigation, launch and flyby trajectory estimation, interacting multiple-model (IMM) filtering, and scientific reconstruction. Finally, the seminar will explore how probabilistic state estimation and modern AI can be combined to develop physics-driven learning algorithms that preserve uncertainty propagation, statistical consistency, and physical constraints.

Past Seminars 2026

July 17 (Fri)

ML Applications in Collider Experiments

#AI4KMI seminar

  • Speaker: Masako Iwasaki (OMU)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • Machine Learning (ML) represents cutting-edge data processing technologies in the field of information science. They are expected to enable various data processing tasks, as they can build a data “model” (a description of the relationship between input and output variables) using information derived from vast amounts of data, even without a pre-existing, explicit model. Since various recent ML techniques provide more effective and precise data processing in accelerator physics experiments, we formed a group with information scientists to apply ML to particle accelerator physics as an RCNP research project in Osaka in 2018. In this seminar, some of our ML application activities in accelerator tuning, physics analysis, and data calibration will be introduced.
July 16 (Thu)

Artificial Intelligence in New Physics Electroweak Phase Transition Studies

#DEEP-IN seminar

  • Speaker: Yang Zhang (Henan Normal University)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • The study of electroweak phase transitions in BSM involves complex numerical calculations, large parameter spaces, and the integration of multiple computational tools. In this talk, I will review recent developments in applying artificial intelligence to new physics phase transition studies. First, I will discuss how machine learning methods can accelerate electroweak phase transition studies, including efficient evaluations of phase transition dynamics, such as bounce action calculations, and the exploration of detectable parameter regions for gravitational-wave searches. Then, I will introduce the emerging role of AI agents in scientific workflows, including automated model construction, effective potential generation, and parameter scans. These developments illustrate how AI can transform traditional computational pipelines and provide new possibilities for future high-energy physics research.

    References

    • Enhancing Phase Transition Calculations with Fitting and Neural Network, arXiv:2510.10667
    • EasyScan_HEP 2: Agent-Ready Parameter Scans for High-Energy Physics, arXiv:2606.31214
July 3 (Fri)

A Cookbook for Collider Metrics: Understanding, Comparing & Combining Event Distances

#CTPU-PTC seminar

  • Speaker: Tianji Cai (Tongji University)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • As particle collider experiments continue to produce ever larger and more complex datasets, a fundamental question arises: how should we measure the similarity between two collider events? A physically meaningful notion of distance lies at the heart of a wide range of applications, from jet tagging to anomaly detection. More fundamentally, it provides a geometric language for collider physics, serving as a common framework for connecting physics-inspired observables with modern machine learning. In this talk, I will present a practical “cookbook” for collider event metrics. Starting from three representative metrics based on optimal transport and relativistic N-body phase space, I will discuss the physical principles encoded by different metrics, how they can be compared on an equal footing, and what aspects of collider events each captures. Finally, I will explore how complementary collider metrics may be combined into a unified framework for event geometry. Beyond providing new tools for collider phenomenology, such a framework offers a principled foundation for understanding, comparing, and designing physics-aware AI models, illustrating how particle physics can serve as a unique testbed for the development of Scientific AI.

  • Tianji Cai (蔡恬吉) is a Distinguished Researcher and Tenure-track Assistant Professor at School of Physical Science and Engineering, Tongji University. Before returning to Shanghai, she was a postdoctoral research associate in the Fundamental Physics Directorate at the SLAC National Accelerator Laboratory, and a research affiliate at the Lawrence Berkeley National Laboratory. She obtained her Ph.D. degree in 2023 at University of California, Santa Barbara, and holds two bachelor's degrees from Duke University and Shanghai Jiao Tong University. Her research explores the interface between High Energy Physics (HEP) and Artificial Intelligence (AI), where she develops first-principles Scientific AI to probe particle phenomenology and, conversely, uses concepts from HEP to deepen our theoretical understanding of AI systems.
July 1 (Wed)

Approximating Euclidean path integrals with radial basis function neural networks

#CTPU-PTC seminar

  • Speaker: Balassa Gabor (Yonsei University)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • In this talk, I will introduce a lattice-based method to approximate Euclidean path integrals, based on a radial basis function (RBF) expansion of the interaction terms that appear in the path integral formalism of quantum field theories. This approach allows numerically efficient determination of both the partition function directly and specific observables, which can be used to describe phenomena such as phase transitions, fluctuations, etc. The method is currently applicable to interacting (real and complex) scalar fields at both zero and non-zero chemical potentials, even in 3+1 dimensions. For real scalar fields in 1+1 dimensions, the phase transition line is approximated at several coupling strengths with very good accuracy, comparable to previous Monte Carlo lattice calculations. As another example, we will examine complex scalar fields at finite chemical potentials, which develop a sign problem similar to that of quantum chromodynamics at finite densities on the lattice. It will be shown that by applying the radial basis expansion to the system, the sign problem can be evaded, and the phase transition points, i.e., the critical chemical potentials where Bose condensation occurs, can be determined. Furthermore, the silver blaze phenomenon, which relies on severe cancellations in the path integral, can also be described. At the end I will propose future directions where the radial basis function approximation could be advantageous, such as systems with fermions, gauge theories, and possibly quantum chromodynamics.

June 10 (Wed)

Parton Showers as a New Path to Light BSM Searches

#CTPU-PTC seminar

  • Speaker: Joon-Bin Lee (Seoul National University)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • Traditional BSM searches focus on new particles produced directly in the hard process. However, light bosons may instead appear through soft or collinear radiation during parton-shower evolution. In this seminar, I will introduce the angular-ordered BSM parton shower implemented in Herwig 7 event generator, and then discuss its phenomenological application to Z′ radiation inside jets. Unlike conventional direct-production searches, shower-induced Z′ production can yield non-isolated dimuon signatures embedded in jets. This motivates a search strategy that is complementary to conventional direct-production searches, while also opening a potentially novel direction based on jet substructure and dimuon-jet correlations.

June 4 (Thu)

DeepQuark: A Deep-Neural-Network Approach to Multiquark Bound States

#DEEP-IN seminar

  • Speaker: Wei-Lin Wu (School of Physics, Peking University)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • Recent discoveries of multiquark candidates have opened a new frontier in hadron spectroscopy and nonperturbative QCD. Understanding these multiquark states poses a challenging quantum many-body problem governed by SU(3) color interactions. Traditional approaches based on basis expansions often encounter severe bottlenecks as the system size and dynamical complexity increase. In this talk, I will present DeepQuark, a deep-neural-network-based variational Monte Carlo framework for solving multiquark bound states. I will discuss the general methodology behind neural-network quantum states, the challenges of extending existing approaches from electronic and nuclear systems to hadron physics, and the architecture of DeepQuark. By combining physics-informed symmetry constructions with the expressive power of deep neural networks, DeepQuark provides a scalable framework for studying multiquark spectroscopy and exploring confinement dynamics.

    References

    • Wei-Lin Wu, Lu Meng, Shi-Lin Zhu, DeepQuark: A Deep-Neural-Network Approach to Multiquark Bound States, Phys.Rev.Lett. 136 7, 071901 (2026), arXiv:2506.20555
June 2 (Tue)

Generative diffusion model with inverse renormalization group flows

#DEEP-IN seminar

  • Speaker: Kanta Masuki (Graduate School of Science, The University of Tokyo)
  • Time: 2:00 PM JST/KST, 1:00 PM Beijing
  • link to the seminar page

  • Diffusion models have recently emerged as one of the most powerful frameworks for generative modeling, achieving remarkable success in a wide range of domains, including image generation, audio synthesis, and scientific data generation. However, despite their empirical success, conventional diffusion models often require many denoising steps and do not explicitly exploit the multiscale structure naturally present in various types of data. This limitation motivates us to ask whether ideas from the renormalization group (RG), which is designed to describe scale-dependent effective degrees of freedom, can provide a useful principle for constructing more efficient generative models.

    In this talk, I will present our recent work on renormalization-group diffusion models (RGDMs) [1], a generative framework that connects diffusion models with RG flows. By establishing a correspondence between diffusion dynamics and exact RG flow equations, we construct a diffusion model whose reverse process generates data in a coarse-to-fine manner, thereby effectively reversing an RG flow.

    I will first introduce the theoretical formulation of RGDMs and explain how the RG perspective leads to a coarse-to-fine generative process. I will then present numerical results in protein structure prediction and image generation, where RGDMs improve sample quality and/or sampling efficiency compared with conventional diffusion models. Finally, I will discuss possible extensions and open questions, including broader applications of RG-inspired generative modeling.

    References

    • K. Masuki and Y. Ashida, Generative diffusion model with inverse renormalization group flows, arXiv:2501.09064
May 20 (Wed)

How AI is changing the way we generate fundamental physics theories and test them in experiments

#PhysAI-Tongji seminar

  • Speaker: Aishik Ghosh (Georgia Tech)
  • Time: 3:30 PM Beijing, 4:30 PM JST/KST
  • With historically some of the biggest datasets in the world, fundamental physics has been one of the early adopters of AI. Today, it is the testbed for uncertainty quantification, density estimation and verification of agentic workflows that also affect the rest of the world. I will discuss how AI has touched every parts of the scientific method and focus on recent developments in hypothesis generation and testing, leveraging the best of both Bayesian and frequentist tools to provide automated but trustworthy scientific results.
  • Aishik Ghosh is an assistant professor at Georgia Institute of Technology, USA. He leads the AI Physicist program supported by The National Energy Research Scientific Computing Center which focusing on the design of AI agents for theoretical physics. His group also designs neural inference techniques for experimental physics and astrophysics, having previously introduced neural simulation-based inference to the ATLAS experiment at CERN. Prof. Ghosh currently contributes to DUNE and the CMS experiment.
May 20 (Wed)

Look-Everywhere Effects in Anomaly Detection

#CTPU-PTC seminar

  • Speaker: Marie Hein (RWTH Aachen)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page

  • To avoid false discoveries in high-energy physics, we typically require a significance exceeding 5σ before claiming a discovery of new physics. Under the background-only hypothesis, the probability of such an excess arising from statistical fluctuations is only about 3×10−7. However, this interpretation is only valid when statistical trials factors are treated correctly. In particular, searches that probe many possible signal configurations — for example by scanning over multiple bins or regions of phase space — are subject to the look-elsewhere effect. Modern model-agnostic searches increasingly rely on machine learning-based anomaly detection, where the search over possible signals is often implicit rather than explicit. This raises important questions about how look-elsewhere effects manifest in these methods. In this seminar, I will compare the statistical behavior of classical binned model-agnostic searches to that of weakly supervised anomaly detection searches. This includes translating between statistics and machine learning terminology to understand parallels and differences in known effects and common practices. Finally, I will discuss how different look-elsewhere effect mitigation strategies impact analysis sensitivity.

Apr. 30 (Thu)

Building autonomous AI physicists for frontier physics research

#DEEP-IN seminar

  • Speaker: Tingjia Miao (School of Artificial Intelligence, Shanghai Jiao Tong University)
  • Time: 1:30 PM JST/KST, 12:30 PM Beijing
  • link to the seminar page

  • Advances in LLMs have led to agents with knowledge and operational capabilities comparable to human scientists, suggesting potential to assist, accelerate, and automate research. Physics, especially theoretical and computational physics, which requires integrating analytical reasoning, code-based computation, and profound domain expertise, is well suited for verifying the end-to-end research capabilities of AI scientists. Accordingly, we construct a general-purpose AI physicist PhysMaster, equipped with a layered academic knowledge base, adapted to the agent skill ecosystem, and adopting an adaptive exploration strategy that balances efficiency and exploration, enabling robust performance in ultra-long-horizon tasks; PhysMaster has been open-sourced. Meanwhile, we introduce PRL-Bench (Physics Research by LLMs), a benchmark with 100 tasks adapted from recent Physical Review Letters papers, covering astrophysics, condensed matter physics, high-energy physics, quantum information, and statistical physics. Evaluation across frontier models shows that failures are dominated by conceptual and formulaic errors, and that exploration and derivations remain unstable over long horizons. In addition, we develop domain-specialized AI scientists, including LQCD Master, which integrates Lattice QCD workflows and expert skills, enabling automated generation and submission of lattice computation scripts from concise physics goals.

    References

    • Miao, Tingjia and others, PhysMaster: Building an Autonomous AI Physicist for Theoretical and Computational Physics Research, arXiv:2512.19799
    • Tan, Jin-Xin, Miao, Ting-Jia and others, Automated Extraction of Collins-Soper Kernel from Lattice QCD using An Autonomous AI Physicist System, arXiv:2603.22471
    • Miao, Tingjia and others, PRL-Bench: A Comprehensive Benchmark Evaluating LLMs’ Capabilities in Frontier Physics Research, arXiv:2604.15411

    Related Links:

Apr. 28 (Tue)

Energy-information trade-off optimizes the cortical critical power law coding

#iPI seminar

  • Speaker: Jun-nosuke Teramae (Kyoto University)
  • Time: 1:30 PM JST/KST, 12:30 PM Beijing
  • link to the seminar page: UTokyo

  • How neurons in the brain represent sensory information is one of the central questions in neuroscience. Recent experiments addressing this problem have revealed that the stimulus responses of cortical neurons exhibit a critical power law. This criticality is hypothesized to balance expressivity and robustness in neural encoding by avoiding the so-called fractal regime, where neural responses become overly sensitive to input perturbations. However, contrary to this assumption, we mathematically prove that neural coding is more robust than previously believed. We develop a theory that provides an analytical expression for the Fisher information in population coding and show that, due to its intrinsic high dimensionality, population coding does not degrade even in the fractal regime. Furthermore, we show that the trade-off between energy consumption and the efficiency of information coding results in the critical power law being the optimal population coding for sensory information.

    Reference: Tatsukawa & Teramae, The cortical critical power law balances energy and information in an optimal fashion, PNAS 122.21, e2418218122 (2025)

Apr. 16 (Thu)

Searching For Anomalies with Foundation Models

#iPI seminar #DEEP-IN seminar

  • Speaker: Vinicius Mikuni (Nagoya University)
  • Time: 2:00 PM JST/KST, 1:00 PM Beijing
  • link to the seminar page: UTokyo
  • link to the seminar page: RIKEN
  • Anomaly detection relaxes the assumptions of how new physics should look and extends the reach of what we can discover. However, interpreting the data and estimating backgrounds remains a challenge. In this new work, we investigate anomalous events selected by the OmniLearned Foundation model across different model sizes, performing a full analysis using CMS Open Data. Surprisingly, models of different sizes, trained on the same data with the same loss functions, select entirely different collisions. In particular, the large OmniLearned model (500M parameters) selects events that are not well described by our background model.
Apr. 15 (Wed)

Toward AI for Physics: From Physical Law Discovery to Scientific Research Agents

#PhysAI-Tongji seminar

  • Speaker: Xiang Li (Peking University)
  • Time: 2:00 PM Beijing, 3:00 PM JST/KST
  • Artificial intelligence in physics should not be viewed merely as a tool for narrow tasks such as formula fitting or benchmark problem solving. More fundamentally, it offers the possibility of enabling AI systems to participate in the formation of physical knowledge and in the actual practice of scientific research. In this talk, I will present our recent efforts toward this broader goal. Using AI-Newton as an example, I will argue that the discovery of universal physical laws requires structured conceptual representations and domain-specific language formulations that go beyond standard symbolic regression. I will then discuss LOCA, which shows how explicit logical frameworks can improve the ability of large language models to perform precise scientific reasoning. Finally, I will introduce Aether, our ongoing effort to develop a more general scientific research agent system equipped with domain expertise, tool-use capabilities, and flexible human-AI interaction. Taken together, these works suggest a promising emerging paradigm for scientific discovery, centered on the integration of knowledge representation, logical reasoning, and research workflows.
  • Xiang Li is a Boya Postdoctoral Fellow at Peking University, working with Prof. Yan-Qing Ma. His research includes AI for science and perturbative quantum field theory. He is particularly interested in how AI can be used to support scientific discovery, with current work spanning AI-driven law discovery, research-oriented AI agent systems, and finding new methods in perturbative QFT. He is actively involved in the development of Aether, an open-source research agent system designed to support scientists in real research workflows. His broader goal is to explore how AI systems can assist human researchers and accelerate progress on frontiers of science.
Apr. 1 (Wed)

Neural Networks from the Perspective of Physics

#CTPU-PTC seminar

  • Speaker: Jaeok Yi (KAIST)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page
  • Despite the remarkable empirical success of deep learning, a comprehensive theoretical understanding of why and how neural networks learn remains a mystery. In this talk, we discuss physics-inspired approaches to understanding neural networks. We present synaptic field theory, a framework that reformulates the gradient descent dynamics of synaptic weights as classical field dynamics in de Sitter spacetime, constructing an action whose metric naturally matches that of a universe with a positive cosmological constant. This framework faces a challenge related to the non-locality of the cost function. To address this issue, we explore the idea of promoting neurons to dynamical degrees of freedom. Leveraging properties of stochastic gradient descent, the Lagrangian can be decomposed into a data-independent bulk part and a data-dependent boundary part. This decomposition is expected to separate the architectural structure from the stochastic properties of neural networks, enabling independent analysis of each. Through this line of research, we aim to provide physicists with a familiar language to investigate the theoretical foundations of machine learning.
Mar. 24 (Tue)

Connecting Simulations and Observations with Differentiable Simulations and Field Level Inference

#AI4KMI seminar

  • Speaker: Benjamin Horowitz (IPMU)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • link to the seminar page
  • The rapid growth of both astrophysical data and simulation capabilities is creating a new opportunity. Instead of being tied to summary statistics (like correlation functions and power spectra), we can begin to connect simulations and observations directly at the field level. In this talk, I will present a framework for field-level, multi-probe inference built around differentiable simulations, where gradients can be propagated through the forward model itself. I will focus on diffhydro, a differentiable hydrodynamics framework written in JAX that combines modern multiphysics solvers with end-to-end automatic differentiation. Starting from simple dark-matter models, we can incrementally add more realistic physics, including turbulence, radiative heating and cooling, and self-gravity, while retaining the ability to optimize directly through the simulation. This makes it possible to connect observations to initial conditions (i.e. latent fields), physical parameters, and unresolved processes in a unified way. I will show how these ideas open the door to reconstructing the history of the Universe and how the same framework can be a platform for embedded machine learning models for additional acceleration and new physics discovery.
Mar. 18 (Wed)

How LLM can help particle physicists

#AI4KMI seminar

  • Speaker: Mihoko Nojiri (KEK)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • link to the seminar page
  • In this talk, I want to discuss evolving field of application of LLM to the scientific coding. The HEP analysis often require lengthy coding of high reliability. We introduce CoLLM, which allows to generate analysis code from the LLM prompts quickly. The package include the automatic bug fixing, and it is now quickly evolving toward code reviews and refinements. I also comments the possible application to the other field and implication from brain functions.
Jan. 29 (Thu)

Status on Generative Unfolding

#AI4KMI seminar

  • Speaker: Sofia Palacios (Rutgers University)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • link to the seminar page
  • Generative machine learning has become a powerful tool for unbinned, high-dimensional unfolding at the LHC. This talk highlights recent progress on key open challenges: scaling to hundreds of dimensions, prior-independent parameter estimation, and the path toward fully analysis-ready unfolding.
Jan. 29 (Thu)

Storage Capacity of Perceptron with Variable Selection

#DEEP-IN seminar #iPI seminar

  • Speaker: Yingying Xu (University of Helsinki)
  • Time: 4:00 PM JST/KST, 3:00 PM Beijing
  • link to the seminar page
  • A central challenge in machine learning is to distinguish genuine structure from chance correlations in high-dimensional data. In this work, we address this issue for the perceptron, a foundational model of neural computation. Specifically, we investigate the relationship between the pattern load α and the variable selection ratio ρ for which a simple perceptron can perfectly classify P = αN random patterns by optimally selecting M = ρN variables out of N variables. While the Cover–Gardner theory establishes that a random subset of ρN dimensions can separate αN random patterns if and only if α < 2ρ, we demonstrate that optimal variable selection can surpass this bound by developing a method, based on the replica method from statistical mechanics, for enumerating the combinations of variables that enable perfect pattern classification. This not only provides a quantitative criterion for distinguishing true structure in the data from spurious regularities, but also yields the storage capacity of associative memory models with sparse asymmetric couplings.
Jan. 29 (Thu)

Physics of Machine Learning

#DEEP-IN seminar #iPI seminar

  • Speaker: Gert Aarts (Swansea University)
  • Time: 2:30 PM JST/KST, 1:30 PM Beijing
  • link to the seminar page
  • In recent years machine learning (ML) has started to make impact in lattice field theory (LFT), e.g. for the generation of ensembles of configurations. In this talk I will explore potential impact in the opposite direction, i.e. using theoretical physics to understand ML approaches. I will relate stochastic gradient descent to random matrix theory and then make the connection between neural networks and disordered systems, leading to a neural network phase diagram in the plane spanned by hyper parameters. I will conclude with the possible impact of our findings for practical ML applications.
Jan. 28 (Wed)

Understanding Galactic Dark Matter with Generative Models

#DEEP-IN seminar #iPI seminar

  • Speaker: Sung Hak Lim (IBS)
  • Time: 2:30 PM JST/KST, 1:30 PM Beijing
  • link to the seminar page
  • Mapping the Milky Way’s dark matter requires moving beyond traditional, rigid dynamical models. In this talk, generative models — specifically Normalizing Flows — are used to learn the stellar phase space distribution directly from Gaia data. This approach enables a flexible, model-independent reconstruction of the Galactic gravitational potential and local dark matter density. These data-driven techniques provide a promising avenue to handle complex observational biases and what they reveal about the dark sector’s influence on our Galaxy.
Jan. 13 (Tue)

Generative AI in Cosmology

#AI4KMI seminar

  • Speaker: Leander Thiele (IPMU)
  • Time: 3:00 PM JST/KST, 2:00 PM Beijing
  • link to the seminar page
  • Increasing data volumes, pushing to non-linear scales, create opportunities for machine learning in cosmology. One primary challenges is the inverse problem implicitly defined through simulations. Neural simulation-based inference is increasingly being recognized as a tool. I will review this technique and present some work both on observational data as well as on methodological development, specifically multi-fidelity inference. In the second part of the talk, I will present recent work on probabilistic identification of cosmic voids.

Past Seminars 2025

Dec. 19 (Fri)

Physics-Driven Learning for Solving Inverse Problems in QCD Physics

#AI4KMI seminar

  • Speaker: Lingxiao Wang (RIKEN)
  • Time: 5:00 PM JST/KST, 4:00 PM Beijing
  • link to the seminar page
  • Discovery in the physical sciences relies on inverse modeling of observations. The combination of deep learning and physics-driven designs is reshaping how we solve inverse problems for extracting physical properties from data. This is particularly relevant for quantum chromodynamics (QCD), where non-trivial symmetries make both data interpretation and computation challenging. In this talk, I will present physics-driven learning from a probabilistic perspective, with a focus on applications in QCD physics. Examples include learning spectral functions and hadron forces from lattice QCD data, reconstructing hadron emission sources from Femtoscopy, and extracting the equation of state from neutron-star observations. If time permits, I will also introduce the physics of diffusion models and discuss physics-driven designs that enable expandable and reliable sampling for accelerating simulations.