Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis

summary

The gist

Frequency-domain thermoreflectance (FDTR) is a laser pump-probe technique widely used to measure thermal properties at the micro- and nanoscale; however…

In short

Vibe-FDTR is an agent framework that lets large language models perform complex frequency-domain thermoreflectance (FDTR) analyses using natural language requests. It combines physical code with agent skills to reliably extract thermal properties from data, achieving high success rates and significantly reducing computational cost compared to traditional methods.

Key concepts

Frequency-domain thermoreflectance (FDTR)
A laser pump-probe technique used at the micro- and nanoscale to measure thermal properties. It involves shining a laser on a material and measuring how it reflects light, which reveals information about heat transport within the material.
Vibe-FDTR Framework Architecture
The core structure consists of two parts: a code layer that enforces physical consistency using configuration files, and a procedural guidance layer that translates user requests into executable analysis steps via agent skills. This separates execution logic from high-level instruction.
Expert Mode
An optional feature where the agent acts as an intelligent copilot for experimental planning. It can autonomously evaluate sensitivity and uncertainty, make physically grounded recommendations for difficult tasks, and suggest complementary measurement modes when data is incomplete.

Terminology used across episodes

This episode discusses

The paper

Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis · Read on arXiv

Fuwei Yang, Weiheng Li, Bai Song

National Key Laboratory of Advanced Micro and Nano Manufacture Technology, Peking University · Center for Nano and Micro Mechanics, Tsinghua University · Department of Engineering Mechanics, Tsinghua University · School of Mechanics and Engineering Science, Peking University

Frequency-domain thermoreflectance (FDTR) is a laser pump-probe technique widely used to measure thermal properties at the micro- and nanoscale; however, it relies on a complex data analysis procedure that demands substantial domain expertise and is susceptible to subtle human errors. Here, we present Vibe-FDTR, an agent-oriented framework that enables large language model (LLM) agents to perform reliable and reproducible FDTR analyses directly from natural language requests. This framework couples a configuration-driven FDTR code package, which enforces physical and parametric consistency, with procedural agent skills that translate user intentions into organized and verifiable analysis steps. We evaluate Vibe-FDTR using a controlled benchmark with two levels: synthetic single-step tasks and real-data multi-step tasks based on measurements of gold-coated graphite samples. Across the two levels, agents using Vibe-FDTR achieve success rates of 100% and 98.9%, respectively. In sharp contrast, ablating skills (Code-agent) reduces performance to 91.4% and 36.7%, which drops further to 38.6% and 0% when the domain package is also omitted (Agent-only). Beyond success rate, Vibe-FDTR also reduces computational cost by 87.7% relative to the Code-agent variant and cuts execution time by more than 60%. Finally, an optional expert mode supports experimental planning via autonomous sensitivity and uncertainty evaluations, and formulates physically grounded recommendations for underspecified tasks. These results demonstrate that encapsulating domain code and expert knowledge into agent skills offers a promising route toward low-barrier, autonomous, and trustworthy thermal metrology.

DOI: 10.1016/j.asi.2026.100049

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis".

Mira: Here's the gist of the paper. Frequency-domain thermoreflectance (FDTR) is a laser pump-probe technique widely used to measure thermal properties at the micro- and nanoscale; however, it relies on a complex data analysis procedure that demands substantial domain expertise and is susceptible to subtle human errors.

Kai: Here, we present Vibe-FDTR, an agent-oriented framework that enables large language model (LLM) agents to perform reliable and reproducible FDTR analyses directly from natural language requests.

Mira: Accurate measurement of thermal properties at the micro- and nanoscale is essential for understanding heat transport and guiding the thermal design of modern electronic and energy systems.

Kai: Despite the widespread application of TDTR and FDTR, a complex and highly demanding data-processing workflow is required to accurately extract the target thermal properties.

Title and authors: Kai: So we're looking at Vibe-FDTR today. This paper introduces a way for large language models to actually do frequency-domain thermoreflectance analysis reliably from just a natural language request.

Mira: It sounds like they are tackling the big headache of FDTR, which is that it needs tons of expert knowledge and you’re prone to making small mistakes with all that data processing.

Lev: Basically, they're building this agent framework that couples a code package enforcing physical rules with these agent skills that turn your plain language into organized steps for the analysis.

Kai: That coupling is key. They are using configuration files as the persistent specifications and then having the procedural guidance layer translate those user requests into executable FDTR operations through specific agent skills.

Mira: It means the AI isn't just guessing at the right fitting parameters or how to organize a complex sequence of decisions; it’s following a defined physical workflow that already has consistency built in.

Lev: And from an engineering standpoint, this is big because it cuts computational cost by eighty-seven point seven percent compared to previous code agent variants and speeds up execution by over sixty percent.

Kai: That efficiency gain is something I really like to hear about, especially when you’re dealing with the massive datasets involved in these kinds of measurements.

Mira: The paper shows that agents using Vibe-FDTR achieve a success rate of one hundred percent for single-step tasks and ninety-eight point nine percent for multi-step tasks using real data.

Lev: That ninety-eight point nine percent on L2 real data is pretty solid, especially when you compare it to the code agent variant which only hit ninety-one point four and thirty six point seven.

Kai: It really shows that encapsulating domain code and expert knowledge into those agent skills offers a way toward low-barrier, autonomous thermal metrology.

Mira: And they have this optional expert mode where the agent can do things like autonomous sensitivity and uncertainty evaluations to plan experiments or suggest physically grounded recommendations when the task is underspecified.

Lev: That ability to recommend complementary measurement modes or identify poorly separable parameters sounds like it could be really useful for someone trying to design an experiment from scratch.

Kai: For design-only tasks, the agent actually assigns explicit physically reasonable assumptions and tests candidate analysis schemes using sensitivity and uncertainty calculations before giving a recommendation.

Mira: It’s essentially acting like an intelligent copilot that knows where the physical model is weak or where your data might be unreliable before you even start fitting.

Lev: That sounds like a significant reduction in the time spent just figuring out the initial analysis plan, which is usually the bottleneck in these kinds of studies.

Kai: So, to summarize Vibe-FDTR: it's an agent-oriented framework that lets LLMs perform reliable FDTR analyses directly from natural language requests by coupling a configuration-driven code package with procedural agent skills.

Mira: It handles the complex, multi-step nature of thermal property extraction by enforcing physical consistency through the code layer and translating user intent into traceable actions via the guidance layer.

Lev: The real win is that it makes this highly demanding analysis accessible to people who might not be deep experts in both cryogenics and advanced data analysis.

Kai: We're seeing improvements where Vibe-FDTR can provide detailed failure mode analysis, pointing out specific errors like bypassing the intended configuration workflow or failing to maintain consistent parameter transfer across stages for easier debugging.

Mira: That level of traceability is what moves this beyond just a successful run to a truly reproducible scientific process.

Lev: It’s about making sure that when you run this on real hardware, you can actually trust where the errors are coming from and fix the underlying skill or configuration.

Kai: So, we've seen how they improve upon previous methods by making the analysis workflow more structured and verifiable through these agent skills.

Mira: The system's ability to handle L2 real-data workflows with a ninety-eight point nine percent success rate really makes this a practical tool for researchers working with actual experimental results.

Lev: It’s moving the goalposts on what we think is possible for autonomous analysis in these complex metrology areas.

Kai: This paper, Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis, really demonstrates how procedural guidance can make complex scientific workflows trustworthy.

Mira: It shows that by giving an agent the right tools and the right way to think about the physics, you lower the barrier for using powerful tools like LLMs in experimental analysis.

Lev: It’s a solid step toward making AI useful not just as a prediction tool, but as an active participant in executing rigorous scientific procedures.

Kai: We'll keep an eye on how this agentic approach develops and see what other complex metrology problems it can tackle next.

The paper's summary: Kai: So, Vibe-FDTR is essentially taking all that messy data analysis in frequency domain thermoreflectance, which used to be this super manual process needing a ton of expert knowledge, and putting it inside an agent framework so large language models can actually do it reliably from plain English.

Mira: Right. It’s about moving away from the kind of one-off fitting where you just throw in parameters hoping for the best, and instead building a whole workflow that checks itself against physics every step of the way.

Lev: I mean, what they’ve built is this layered system with a code layer that handles all the physical rules—like making sure your thermal model doesn't say something physically impossible—and then an agent skills layer that takes what you ask for in words and translates it into those specific, verifiable analysis steps.

Kai: That procedural guidance layer is what I’m most interested in. It means the agent isn't just guessing; it’s following a pre-defined path for the analysis, which is huge when you're dealing with something as non-linear as thermal transport at the microscale.

Mira: Exactly. The paper shows that this approach achieves very high success rates on tasks involving real experimental data, specifically ninety-eight point nine percent for multi-step analyses, which is a big deal compared to what we used to see with simpler agent setups.

Lev: And it’s not just about the final number; they show how it cuts down on the computational cost by eighty-seven point seven percent compared to running just the code agent, and execution time drops by over sixty percent. That matters because running these complex analyses can be really time-consuming on standard hardware.

Kai: It makes sense that efficiency is a big part of this for me, especially since my work with cryogenics and qubits involves managing huge amounts of data where speed and resource usage are always concerns.

Mira: And they have this optional expert mode, which lets the agent proactively check its own work by doing sensitivity evaluations and uncertainty assessments before suggesting an analysis plan. It’s like having an assistant that can tell you if your experimental setup is underspecified or if a certain parameter is just too noisy to trust right now.

Lev: For design tasks, it’s even better because the agent can assign reasonable assumptions and test different ways of analyzing the data using those sensitivity calculations before recommending a final scheme. It’s testing hypotheses autonomously.

Kai: It really shows that encapsulating deep domain knowledge and validated physical code into agent skills is a viable path toward making this kind of thermal metrology accessible to people who aren't just specialists in the field anymore.

Mira: The real value here is how it lowers the barrier for entry. You don't need to be an expert in cryogenics *and* a master of fitting routines; you just need to know how to ask the right question, and Vibe-FDTR handles the rest by guiding the process through physical constraints.

Lev: It’s about taking that invaluable human-expert experience—the "know-how"—and baking it into skills so that a new practitioner or someone from a different field can use powerful tools like language models to get reproducible results.

Kai: So, while we’re looking at how this handles complex thermal properties, I wonder what the limitations are for things outside of just FDTR data fitting. Does this framework have trouble with entirely different types of physical problems?

Mira: That’s a fair point. The paper does mention that the method is specifically tuned for the structure of thermoreflectance data analysis, so it really excels where that physics applies, but it’s not a universal solver for every kind of measurement problem out there.

Lev: And you have to remember, even with all this agent guidance and code checks, if the underlying physical model itself is flawed or incomplete, the output will still be based on that flawed foundation.

Kai: That makes sense. So, it’s a very powerful tool for FDTR analysis right now—reliable and efficient—but we've got to keep an eye on how these agent skills can generalize when we try to apply this framework to entirely different types of physical measurements in the future.

The paper's improvements: Kai: So, we’re looking at what they suggest for improvements to Vibe-FDTR today: they are talking about how to make this agent framework even more robust and useful moving forward than what they presented in the paper.

Mira: They want to focus on making the procedural guidance layer smarter so that it can handle even messier user requests without needing a perfect, pre-written instruction for every single analysis.

Lev: I think they’re pushing for better ways to make those agent skills more adaptable, meaning they should be able to handle a wider variety of underlying physical models or different measurement setups than just the ones tested in the benchmark tasks.

Kai: That makes sense. Right now, it’s great on the benchmark stuff, but I need to know if this framework can actually handle novel experimental data without needing a complete rewrite of the agent's skill set.

Mira: Exactly. They suggest incorporating better mechanisms for self-correction within the agent itself, so when it hits a weird data pattern, it doesn't just fail; it should be able to identify *why* it failed and try a different approach guided by those expert evaluations we talked about earlier.

Lev: From my side, I think they’re focusing on making the code layer even more modular so that if you have a very specific, custom physical model—say, one for a new material—you can plug in that custom physics without having to overhaul the whole framework structure.

Kai: That’s important because my work often involves tailoring models to specific superconducting materials, so being able to swap in those custom physics efficiently is key for real hardware testing.

Mira: And they also seem keen on developing more sophisticated ways for the agent to communicate its reasoning back to the user, not just showing the final result, but explaining exactly which physical assumptions it made at each decision point.

Lev: That level of detailed introspection is crucial because when you’re debugging a complex thermal measurement, knowing precisely where the AI decided to make a specific choice helps you pinpoint if the error came from an initial assumption or from a bad calculation step later on.

Kai: It sounds like they are moving toward a more transparent system where we can really audit the decision-making process of the agent during an analysis run.

Mira: Right. They want to ensure that this tool doesn't become a black box where you just trust the output without knowing the underlying physical reasoning behind those ninety-eight point nine percent success rates.

Lev: It’s about building in more explicit checks for consistency across stages, making sure parameter transfers are tracked meticulously from the start of the analysis all the way to the final output.

Kai: So, if you take it all together, they're not just looking at getting better numbers on FDTR; they’re looking at making this agent framework a more flexible and transparent scientific partner for experimentalists.

Mira: And that flexibility is what will likely allow researchers to push the boundaries of what kind of complex thermal data analysis can be automated and trusted.

Conclusion: Kai: So we’re wrapping up on Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis, which is essentially this new way to let language models perform complex thermal property measurements reliably from just natural language requests.

Mira: It’s a really solid piece of work because it takes the incredibly demanding and often manual process of FDTR analysis and wraps it into a verifiable, agent-driven workflow that enforces physical rules.

Lev: And the efficiency gains are notable; cutting computational cost by eighty-seven point seven percent is significant when you think about running these kinds of simulations on real hardware, which usually requires a lot of overhead.

Kai: I agree. It shows that we can get more complex analysis done faster and with higher fidelity than relying on just simple fitting operations or basic code agents.

Mira: The big implication for me is how this democratizes access to this type of measurement; it lowers the barrier so people who aren't thermal metrology experts can use these powerful techniques without needing years of specialized training.

Lev: I just want to say that the success rates they hit on real data, like ninety-eight point nine percent for multi-step tasks, really tells you that this isn't just theoretical potential; it’s actually working on the kind of data you’d get in a lab setting.

Kai: It makes me think about where this can go next. If we can make these agents so good at handling thermal data analysis, what other kinds of complex physical measurements could an agent framework like this tackle?

Mira: That’s where the future lies; if we can prove this approach works across different physical domains, it opens up the door to applying this structured, agent-guided procedural logic to other areas of condensed matter physics or even materials science.

Lev: I just want to stress that for any of this agent work to be truly useful on real hardware, the underlying code and the skills have to be extremely stable and reproducible, which is what they’ve focused on with those configuration files.

Kai: So we’re looking at a system that's not only fast and accurate but also built with a strong foundation for future scaling.

Mira: Exactly. The goal here is to create tools that are trustworthy because they can explain their reasoning through the physics, not just spit out an answer in some black box format.

Lev: It’s about building an experimental assistant that doesn't just suggest parameters but actually guides the entire experimental plan with uncertainty checks built in from the start.

Kai: So we’ve seen how Vibe-FDTR can streamline FDTR analysis, and it points toward a future where agent frameworks are central to complex scientific data processing.

More episodes

← Home