PRISM: Unifying multiphysics optothermal transport and identifiability-guided inference for nanoscale thermal metrology
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Nonlinear Sciences. The day in nonlinear science: chaos, patterns, networks and adaptation.
Kai: We're looking at the paper: "PRISM: Unifying multiphysics optothermal transport and identifiability-guided inference for nanoscale thermal metrology".
Mira: Here's the gist of the paper. PRISM (Physics-Resolved Inference and Screening Meta-framework) is an agentic scientific architecture designed to unify a hierarchy of optothermal transport solvers with automated model routing and identifiability screening to resolve ill-posed multi-parameter inversions in micro- and nanoscale thermal metrology.
Kai: The PRISM forward model hierarchy spans four dimensions: optical heating/detection (SE), electron–phonon coupling (SX), phonon transport (ST), and spatial depth parameterization (SR).
Mira: It progresses logically from the diffusive surface-heating, single-temperature model (SH-STM) to fully coupled nonFourier formulations: SH-STM to VH-STM to OTh-STM to OTh-TTM to OTh-TTM(NF).
Kai: Diffusive layers utilize STM, metallic films employ TTM to resolve electron and phonon temperatures, and nonmetallic layers are solved via STM with directional thermal conductivities.
Paper summary: Kai: The paper PRISM introduces this forward model hierarchy spanning four dimensions, moving from simple diffusive surface heating to more complex models like OTh-TTM(NF) for non-Fourier substrate closure.
Mira: They show how they use this structure to handle different material types, like using the single temperature model for diffusive layers and TTM for metallic films.
Lev: And the key is that they use a stateful LangGraph execution graph that alternates between reasoning, which involves an LLM, and acting by dispatching tools or terminating based on physical scaling criteria.
Kai: This agent maintains a stateful context including data quality, Jacobian matrices, singular value decomposition results, and the fit history throughout the process.
Mira: They use a structured domain knowledge graph that acts like a wiki of cited literature and solver code to guide model revisions when they find discrepancies in the fit.
Lev: The screening logic involves data diagnosis followed by sensitivity estimation using logarithmic sensitivity Jacobian analysis, or Sj.
Kai: Then they use noise-whitened singular value decomposition, or SVD, to construct the retained singular mode subspace when the ratio of the singular values is greater than zero point one.
Mira: And they use a planner to filter your requested targets by checking if they are in the final ordered intersection list derived from combining sensitivity and identifiability lists.
Lev: They expose a set of tools like diagnose data, estimate sensitivity, analyze identifiability, compute phase, and fit parameters for the optimization step.
Kai: Optimization is done via reduced-space global optimization using an optimizer like DTDE because it's efficient for this kind of problem.
Mira: And they also have a specific toolset that autonomously constructs and verifies a Deep Operator Network solver when dealing with ion-irradiated or graded media where the conductivity profile k(z) is continuous.
Lev: This DeepONet reconstruction helps get steep damage gradients and significantly reduces the root mean square percentage error, or RMSPE.
Kai: They demonstrate this by resolving cross-regime parameter distortions, like correcting a twenty-four to twenty-eight percent interfacial conductance underestimation in gold silicon stacks by decoupling the near-interface electron phonon equilibration resistance.
Mira: And in bimetallic gold titanium silicon contacts, they use nonlocal modeling to reconcile scatter issues into an intrinsic silicon conductivity of one hundred forty point zero five to one hundred forty point one four watts per meter kelvin.
Lev: The framework is designed to be extensible across different thermalwave techniques like TDTR or three omega, allowing reuse of the layerwise TTM and non-Fourier models with technique-specific source, readout, and noise models.
Conclusion: Kai: So PRISM is this agentic architecture that unifies transport solvers with automated routing to solve ill-posed problems in nanoscale thermal metrology.
Mira: The title itself suggests a unification of multiphysics optothermal transport with identifiability-guided inference across the whole system.
Lev: What this means for the hardware guys is a way to run these complex fits on real hardware by using reduced space global optimization efficiently, which is what DTDE does well.
Kai: For the experimentalists, it means they can use this framework to diagnose where their data fitting goes wrong by looking at sensitivity and identifiability lists before they waste time trying every parameter combination.
Mira: The implication is that we can tackle those real-world material problems, like those interfacial conductance issues in Au silicon stacks, with a much more robust inference method than traditional fitting routines.
Lev: It shows how to manage the complexity of coupling electronic and phononic channels when you have continuous conductivity profiles k(z) by building that DeepONet solver autonomously.
Kai: The authors are showing that by systematically diagnosing parameter dependencies and identifying the relevant subspace, we can achieve high fidelity estimates even when the input data is inherently noisy or limited.
Mira: It’s about moving away from just fitting curves to using an agent that understands the underlying physics hierarchy to intelligently select which physical models apply at any given scale.
Yicong QIU, Ziyang ZHONG, Ziyan QIAN, Yang LU, Yong Jian HONG, Chenhan LIU, Qiye ZHENG
Department of Mechanical and Aerospace Engineering, The Hong Kong University of Science and Technology · Department of Computer Science and Engineering, The Hong Kong University of Science and Technology · Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology · School of Energy and Mechanical Engineering, Nanjing Normal University
nlin.CD
Submitted: 2026-11-02
Updated: 2026-11-02
Comments: Includes Supplemental Information
Code: https://github.com/langchain-ai/langgraph
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 100/100
The gist: PRISM (Physics-Resolved Inference and Screening Meta-framework) is an agentic scientific architecture designed to unify a hierarchy of optothermal transport solvers with automated model routing and…
Key concepts
- Forward Model Hierarchy
- PRISM uses a structured sequence of thermal models, starting from simple diffusive surface heating up to fully coupled nonFourier formulations. This hierarchy ensures that the correct level of physical detail is applied depending on the material layer being analyzed, allowing the system to handle diverse thermal transport physics systematically.
- Identifiability Screening
- This logic uses mathematical tools like Singular Value Decomposition (SVD) on noise-whitened data to determine which parameters in a complex model can actually be reliably measured. It filters out unidentifiable parameters, ensuring the final results are physically meaningful and robust against measurement noise.
- Model Routing
- The agent automatically chooses the most appropriate thermal solver for a given physical scenario based on dimensionless scaling criteria. This intelligent routing prevents applying an overly complex model when a simpler one suffices, optimizing computational efficiency while maintaining accuracy.
- Deep Operator Network (DeepONet)
- For materials with continuous conductivity profiles, PRISM can automatically build and verify specialized neural network solvers. These DeepONets reconstruct steep damage gradients accurately, significantly reducing the error in measurements that rely on complex spatial variations.
Terminology
Summary
PRISM (Physics-Resolved Inference and Screening Meta-framework) is an agentic scientific architecture designed to unify a hierarchy of optothermal transport solvers with automated model routing and identifiability screening to resolve ill-posed multi-parameter inversions in micro- and nanoscale thermal metrology. p. 3
The PRISM forward model hierarchy spans four dimensions: optical heating/detection (SE), electron–phonon coupling (SX), phonon transport (ST), and spatial depth parameterization (SR). p. 5 It progresses logically from the diffusive surface-heating, single-temperature model (SH-STM) to fully coupled nonFourier formulations: SH-STM to VH-STM to OTh-STM to OTh-TTM to OTh-TTM(NF). p. 4 Diffusive layers utilize STM, metallic films employ TTM to resolve electron and phonon temperatures, and nonmetallic layers are solved via STM with directional thermal conductivities. p. 5 The nonFourier substrate closure, OTh-TTM(NF), derived from the Guyer–Krumhansl relation, is invoked when diffusive hierarchy fails. p. 4
The PRISM execution layer is implemented as a stateful LangGraph execution graph that alternates between reasoning (invoking an LLM) and acting (dispatching tools or terminating). p. 3 It parses tasks to route branches based on dimensionless physical scaling criteria, maintains a stateful context including data quality, Jacobian, SVD results, and fit history. p. 3 The agent utilizes a structured domain knowledge graph—a wiki of cited literature and solver code—for model revision guided by misfit diagnosis. p. 5
The PRISM screening logic involves data diagnosis, sensitivity estimation via logarithmic sensitivity Jacobian analysis (Sj), identifiability analysis using noise-whitened Singular Value Decomposition (SVD) to construct the retained singular-mode subspace (ISVD = sigma m: sigma m/sigma 1 > 0. p. 31), and target reduction. p. 1 The agent employs a planner to filter requested targets by membership in the final ordered intersection list, which is derived from combining sensitivity and identifiability lists (Lident). p. 9
The PRISM solver toolset exposes functions such as diagnose_data, estimate_sensitivity, analyze_identifiability, compute_phase, and fit_parameters. Optimization is performed via reduced-space global optimization, typically using the DTDE optimizer due to its efficiency. p. 3 For ion-irradiated or graded media with continuous conductivity profiles k(z), PRISM autonomously constructs and verifies a Deep Operator Network (DeepONet) solver to reconstruct steep damage gradients, achieving significant reduction in root-mean-square percentage error (RMSPE). p. 3
The framework demonstrates its capability by resolving pervasive cross-regime parameter distortions. p. 3 For instance, in Au/Si stacks, it corrects a 24–28% interfacial conductance underestimation by decoupling the near-interface electron–phonon equilibration resistance. p. 0 In bimetallic Au/Ti/Si contacts, nonlocal modeling reconciles severe spot-size and frequency-window scatter into a configuration-independent intrinsic silicon conductivity of 140. p. 35--141.4 W m-1 K-1. The PRISM framework is extensible across various thermalwave techniques (e. p. 0g., TDTR, 3 omega) and can be adapted for different probing modalities by reusing the layerwise TTM and non-Fourier models while incorporating technique-specific source, readout, and noise models. p. 2
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Agentic framework integration for automated model selection: The PRISM agent can "evaluate dimensionless physical scaling criteria from the domain knowledge graph—the optical penetration ratio dtrans/dopt, carrier equilibration ratio dtrans/Lep,z, and Knudsen number Kn = λph/ min(wr, δp)—alongside forward-phase deviation gates" to assemble the required solver branch.
-
SVD guided parameter screening: The system can
prune unidentifiable parameters through noise-whitened singular value decomposition (SVD)
by retaining only those parameters where "σm/σ1 > τSNR," thereby ensuring that only physically relevant targets are included in global optimization. -
Autonomous solver construction for continuous profiles: For functionally graded materials, the system can
autonomously constructs and verifies a deep operator network (DeepONet) solver
to reconstruct k(z), avoiding the need to fitdozens of thin sublayers
and reducing error from 3.3% to 2.7%. -
Cross-configuration validation: The agent can perform
post-fit cross-configuration consistency checks
across different measurement setups, such as coaxial versus offset beam geometries, to ensureconvergence of kSi to 141 W m–1 K–1 across all six spot-size and frequency-window combinations in Au/Ti/Si.
-
Operator learning for continuous profile reconstruction: The DeepONet architecture can be used to reconstruct continuous conductivity profiles by encoding the input phase spectrum into a latent vector b and mapping query depth z to basis functions t(z) to synthesize the profile bk(z) = b · t(z) + b0.
Related papers
- Quantifying The Complex Spatiotemporal Chaos of Cardiac Fibrillation in Ionic Models Across Parameter Regimes
- Modal interactions in strongly nonlinear coupled oscillators with closely spaced modes and nonclassical damping distribution
- An active Lorentz gas: walking droplets in periodic media
- Stochastic-layer width as an order parameter for the integrability-to-chaos transition in the bouncer model
- When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates
- Universal Scaling Laws in Freeway Traffic