arxiv: 2604.21220 · v1 · submitted 2026-04-23 · ⚛️ physics.app-ph

Recognition: unknown

A transfer-learning-enhanced POD-FNN surrogate for rapid signal prediction and inverse fitting in thermoreflectance with patterned transducers

Bingjia Xiao , Tao Chen , Puqing Jiang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 13:18 UTC · model grok-4.3

classification ⚛️ physics.app-ph

keywords thermoreflectancesurrogate modelingproper orthogonal decompositiontransfer learninginverse fittingpatterned transducersthermal conductivity

0 comments

The pith

A POD-FNN surrogate with transfer learning predicts thermoreflectance signals 534 times faster while preserving accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that proper orthogonal decomposition of phase signals followed by a feedforward neural network can replace repeated high-fidelity simulations in patterned-transducer thermoreflectance. Training the network on parameter-to-coefficient mappings, then applying transfer learning for new domains, produces mean RMSE below 0.2 degrees and cuts prediction time from 5.39 s to 0.01 s per signal. The same surrogate reduces a representative inverse fit from roughly 18950 s to 65 s, and yields stable silica thermal conductivities when run on measured Al/SiO2 data.

Core claim

Within the original parameter domain, the surrogate achieves mean and median RMSE values of 0.19 and 0.17 degrees, with a maximum RMSE below 0.47 degrees, while reducing the average prediction time per signal from 5.39 s to 0.01 s (about 534x). In inverse analysis, the fitting time for a representative case is reduced from about 18950 s to about 65 s with comparable accuracy. Transfer learning further improves performance in expanded parameter domains, with the TL-FR strategy giving the best overall results.

What carries the argument

A feedforward neural network that maps thermophysical and geometric parameters to the coefficients of a proper-orthogonal-decomposition basis built directly from simulated phase signals, with transfer learning used to adapt the network when the parameter domain expands.

If this is right

Iterative inverse fitting for multiple transducer patterns becomes practical instead of computationally prohibitive.
High-fidelity data generation for new domains drops from roughly 34000 s to under 6000 s when target samples are reduced to 1000.
Stable material-property extraction is demonstrated on real Al/SiO2 samples across conventional and patterned geometries.
The surrogate supports repeated forward evaluations needed for model updating in thermoreflectance workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same POD-FNN structure could be reused for other time-resolved heat-transport measurements that currently rely on slow finite-element runs.
Embedding the surrogate inside measurement software could enable on-the-fly adjustment of laser-spot sizes or modulation frequencies.
Further tests with target datasets smaller than 1000 samples would clarify the minimum high-fidelity data still required for reliable transfer.

Load-bearing premise

The validated COMSOL model accurately represents the physical thermoreflectance response across the full range of thermophysical and geometric parameters needed for both training and target domains.

What would settle it

New experimental phase signals or independently measured thermal conductivities that deviate systematically from the surrogate predictions and fitted values beyond the reported RMSE levels.

read the original abstract

Patterned-transducer thermoreflectance enhances sensitivity to low-thermal-conductivity materials by suppressing lateral heat spreading in the metal transducer, but its wider use is limited by the cost of repeated high-fidelity forward evaluations in iterative fitting. Here, we develop a transfer-learning-enhanced POD-FNN surrogate for rapid phase prediction in patterned-transducer thermoreflectance, using patterned FDTR as a representative case. A validated COMSOL model is first constructed, and proper orthogonal decomposition is applied directly to the phase signals to build a compact reduced-order representation. A feedforward neural network is then trained to predict the POD coefficients from thermophysical and geometric parameters. Within the original parameter domain, the surrogate achieves mean and median RMSE values of 0.19 and 0.17 degrees, with a maximum RMSE below 0.47 degrees, while reducing the average prediction time per signal from 5.39 s to 0.01 s (about 534x). In inverse analysis, the fitting time for a representative case is reduced from about 18950 s to about 65 s with comparable accuracy. The framework is further applied to measured Al/SiO2 samples, yielding stable silica thermal conductivities of 1.44 +/- 0.088, 1.43 +/- 0.093, and 1.50 +/- 0.079 W/(m K) for conventional FDTR and patterned FDTR with pattern radii of 5.3 and 3.25 um, respectively. Transfer learning further improves performance in expanded parameter domains, with the TL-FR strategy giving the best overall results. Reducing the additional target-domain dataset from 6000 to 1000 samples also lowers the high-fidelity data-generation time from about 34179 s to about 5885 s. The proposed framework provides an accurate and efficient route for repeated forward evaluation, rapid inverse fitting, and cost-effective model updating in patterned thermoreflectance workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a POD-FNN surrogate with transfer learning that cuts phase prediction time by 534x and fitting time by hundreds of times for patterned FDTR, but all accuracy numbers are measured only against the COMSOL model it was trained on.

read the letter

The main takeaway is that this surrogate delivers large practical speedups for repeated forward evaluations and inverse fitting in patterned-transducer thermoreflectance. Within the training domain it reports mean and median phase RMSE of 0.19 and 0.17 degrees with a max under 0.47, drops average prediction time from 5.39 s to 0.01 s, and reduces a sample inverse fit from roughly 18950 s to 65 s while producing stable silica conductivity values around 1.44–1.50 W m⁻¹ K⁻¹ on measured Al/SiO2 samples. Transfer learning lets them cut the new high-fidelity data needed for expanded domains from 6000 to 1000 samples. Those are the concrete results that matter for anyone running many fits on thin-film thermal data. The workflow itself is straightforward: validated COMSOL runs generate phase signals, POD compresses them, an FNN maps parameters to coefficients, and transfer learning adapts the network with limited extra data. The paper shows the numbers clearly and demonstrates the end-to-end use on real samples, which is the useful part. The soft spot is exactly what the stress-test note flags. Every quantitative accuracy claim, including the “comparable accuracy” statement for inverse fitting, comes from comparing the surrogate to the same COMSOL model used for training. The experimental section only reports that the fitted k values stay consistent across transducer radii; it does not show residuals between the surrogate-predicted phases and the actual measured phase signals. If the COMSOL model has any systematic offset from real thermoreflectance physics, that offset passes straight into the surrogate without being visible in the reported metrics. This is a limitation, not a fatal flaw, but it is the central one. The work is for people who already do frequency-domain thermoreflectance on patterned structures and need faster iteration for parameter sweeps or fitting. It is not a new theoretical framework, just a competent implementation of existing POD-plus-network ideas in this niche. I would send it to peer review because the speedups are large enough to be worth documenting and the experimental consistency check is present, even though reviewers will almost certainly request direct surrogate-to-measurement phase comparisons.

Referee Report

2 major / 2 minor

Summary. The manuscript develops a transfer-learning-enhanced POD-FNN surrogate for rapid phase-signal prediction and inverse parameter fitting in patterned-transducer frequency-domain thermoreflectance (FDTR). A validated COMSOL model generates training data; POD reduces the signals to coefficients that a feedforward network maps from thermophysical and geometric inputs. Within the training domain the surrogate reports mean/median/max RMSE of 0.19/0.17/<0.47°, a 534× speedup versus COMSOL, and fitting-time reduction from ~18950 s to ~65 s with comparable accuracy. Transfer learning is shown to extend the domain with reduced target data, and the framework is applied to experimental Al/SiO2 samples, producing stable silica conductivities of 1.44–1.50 W m⁻¹ K⁻¹ across transducer radii.

Significance. If the surrogate’s accuracy against COMSOL translates to measured signals, the method would remove the dominant computational cost of iterative fitting in patterned thermoreflectance, enabling denser sampling of parameter space and routine use of complex transducer geometries. The explicit reporting of RMSE statistics, wall-clock timings, and transfer-learning data-reduction factors (6000→1000 samples) provides concrete, reproducible evidence of the claimed efficiency gains.

major comments (2)

[Abstract and inverse-analysis results] Abstract and §4 (inverse-analysis results): all quantitative accuracy claims (RMSE 0.19/0.17/<0.47°, “comparable accuracy” in fitting) are computed exclusively against the COMSOL forward model. No residuals between surrogate-predicted phase and the actual measured experimental phase signals are reported, so any systematic discrepancy between the COMSOL model and real thermoreflectance physics (interface resistance, transducer roughness, laser-spot effects) would propagate undetected into the fitted conductivities.
[Transfer-learning experiments] §3.3 and transfer-learning experiments: the claim that TL-FR yields the best overall results is supported only by synthetic-target-domain metrics; it is unclear whether the same ranking holds when the target domain includes experimental noise or when the reduced 1000-sample set is drawn from the same distribution as the 6000-sample set.

minor comments (2)

[Methods] Notation for the POD truncation rank and the precise definition of the phase RMSE (absolute or relative, per-frequency or integrated) should be stated explicitly in the methods section.
[Experimental results] Figure captions for the experimental fitting results should include the number of frequency points and the frequency range used in the least-squares objective.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us identify areas where the manuscript can be strengthened. We address each major comment point by point below and outline the revisions we will make to the manuscript.

read point-by-point responses

Referee: [Abstract and inverse-analysis results] Abstract and §4 (inverse-analysis results): all quantitative accuracy claims (RMSE 0.19/0.17/<0.47°, “comparable accuracy” in fitting) are computed exclusively against the COMSOL forward model. No residuals between surrogate-predicted phase and the actual measured experimental phase signals are reported, so any systematic discrepancy between the COMSOL model and real thermoreflectance physics (interface resistance, transducer roughness, laser-spot effects) would propagate undetected into the fitted conductivities.

Authors: We agree that the reported RMSE statistics and fitting accuracy are evaluated against the COMSOL forward model, as the surrogate is designed to emulate it. The COMSOL model was previously validated against experimental data (as noted in the manuscript), and the inverse results on measured Al/SiO2 samples produce consistent silica conductivities across transducer geometries, providing indirect support. However, we acknowledge the value of direct residuals. In the revised manuscript we will add a new panel to Figure 8 (or a dedicated supplementary figure) that overlays surrogate-predicted phase signals (using the fitted parameters) against the raw experimental phase data for all three transducer radii, together with the corresponding residuals. This addition will quantify any residual model-experiment mismatch and confirm that it remains small relative to the fitting uncertainty. revision: yes
Referee: [Transfer-learning experiments] §3.3 and transfer-learning experiments: the claim that TL-FR yields the best overall results is supported only by synthetic-target-domain metrics; it is unclear whether the same ranking holds when the target domain includes experimental noise or when the reduced 1000-sample set is drawn from the same distribution as the 6000-sample set.

Authors: The 1000-sample target set is drawn from the identical parameter distribution as the 6000-sample set, as stated in §3.3 and the methods. The TL ranking is therefore evaluated on data from the same distribution. Regarding experimental noise, we recognize that the current tests use noise-free synthetic signals. To address this, we will include an additional robustness test in the revised §3.3: Gaussian noise with standard deviation matching the typical phase noise level observed in our FDTR experiments (~0.1–0.2°) will be added to the target-domain signals, and the performance of TL-FR, TL-FT, and other strategies will be re-ranked. We expect TL-FR to remain superior, but the new results will be reported explicitly so readers can judge the claim under realistic noise conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; surrogate is a learned mapping validated against independent simulation and experiment

full rationale

The paper's core chain begins with an independently validated COMSOL finite-element model used to generate synthetic phase signals, followed by POD to obtain a reduced basis and training of an FNN to map thermophysical/geometric inputs to POD coefficients. Reported RMSE values (mean 0.19°, median 0.17°, max <0.47°) are computed on held-out COMSOL data that is statistically independent of the training set; the inverse-fitting speed-up is likewise measured against direct COMSOL optimization. Application to measured Al/SiO2 signals produces stable k values across transducer radii without any algebraic identity that would force surrogate outputs to equal training targets by construction. No self-definitional step, fitted-input-called-prediction, or load-bearing self-citation appears in the derivation; the model is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the fidelity of the COMSOL finite-element model for generating training data and on the generalization ability of the POD-FNN across parameter domains via transfer learning. No new physical entities are postulated.

free parameters (2)

POD truncation rank
Number of modes retained to represent phase signals; chosen to balance compactness and accuracy but value not stated in abstract.
Neural network architecture and training hyperparameters
Layer sizes, learning rates, and transfer-learning strategy parameters fitted during training on simulation data.

axioms (2)

domain assumption The COMSOL finite-element model accurately captures the thermoreflectance physics for the patterned transducer geometry.
Used to generate all training and validation signals.
standard math Proper orthogonal decomposition yields an effective low-dimensional basis for the phase signals.
Standard linear algebra technique for signal compression.

pith-pipeline@v0.9.0 · 5672 in / 1676 out tokens · 54125 ms · 2026-05-08T13:18:11.306443+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references

[1]

Mohan, S

R. Mohan, S. Khan, R.B. Wilson, P.E. Hopkins, Time -domain thermoreflectance, Nature Reviews Methods Primers, 5 (2025) 55

2025
[2]

Jiang, X

P. Jiang, X. Qian, R. Yang, Tutorial: Time -domain thermoreflectance (TDTR) for thermal property characterization of bulk and thin film materials, Journal of Applied Physics, 124 (2018)

2018
[3]

T. Chen, S. Song, Y . Shen, K. Zhang, P. Jiang, Simultaneous measurement of thermal conductivity and heat capacity across diverse materials using the square-pulsed source (SPS) technique, International Communications in Heat and Mass Transfer, 158 (2024) 107849

2024
[4]

Kirsch, J

D.J. Kirsch, J. Martin, R. Warzoha, M. McLean, D. Windover, I. Takeuchi, An instrumentation guide to measuring thermal conductivity using frequency domain thermoreflectance (FDTR), Review of Scientific Instruments, 95 (2024)

2024
[5]

L. Tang, C. Dames, Anisotropic thermal conductivity tensor measurements using beam-offset frequency domain thermoreflectance (BO-FDTR) for materials lacking in- plane symmetry, International Journal of Heat Mass Transfer, 164 (2021) 120600

2021
[6]

Jiang, D

P. Jiang, D. Wang, Z. Xiang, R. Yang, H. Ban, A new spatial -domain thermoreflectance method to measure a broad range of anisotropic in -plane thermal conductivity, International Journal of Heat Mass Transfer, 191 (2022) 122849

2022
[7]

M. Goni, M. Patelka, S. Ikeda, T. Sato, A.J. Schmidt, Frequency domain thermoreflectance technique for measuring the thermal conductivity of individual micro-particles, Review of Scientific Instruments, 89 (2018)

2018
[8]

Akura, Y

Y . Akura, Y . Matsunaga, L. Liu, Y . Ikeda, M. Shimofuri, A. Banerjee, T. Tsuchiya, J. Hirotani, Enhancing the precision of thermal conductivity measurement via transducer patterning in frequency -domain thermoreflectance, Review of Scientific Instruments, 96 (2025)

2025
[9]

Warzoha, A.A

R.J. Warzoha, A.A. Wilson, B.F. Donovan, A.N. Smith, N. Vu, T. Perry, L. Li, N. Miljkovic, E. Getto, A numerical fitting routine for frequency -domain thermoreflectance measurements of nanoscale material systems having arbitrary geometries, Journal of Applied Physics, 129 (2021)

2021
[10]

Warzoha, A.A

R.J. Warzoha, A.A. Wilson, B.F. Donovan, A. Clark, X. Cheng, L. An, G. Feng, Measurements of thermal resistance across buried interfaces with frequency -domain thermoreflectance and microscale confinement, ACS Applied Materials Interfaces, 16 (2024) 41633-41641

2024
[11]

X.-Y . Feng, F. Bai, W.-Q. Tao, A new efficient conservation -based method for implementing pod -Galerkin projection, in: International Heat Transfer Conference Digital Library, Begel House Inc., 2023

2023
[12]

Xiang, B

L. Xiang, B. Zhang, Y . Zha, G. Xing, X. Yang, Z. Wang, Y . Cheng, X. Yu, R. Hu, X. Luo, Physics-informed proper orthogonal decomposition for accurate and superfast prediction of thermal field, ASME Journal of Heat Mass Transfer, 147 (2025) 073301

2025
[13]

Allabou, R

M. Allabou, R. Bouclier, P.-A. Garambois, J. Monnier, Reduction of the shallow water system by an error aware POD-neural network method: Application to floodplain dynamics, Computer Methods in Applied Mechanics Engineering, 428 (2024) 117094

2024
[14]

Y .-Q. Tang, W. -Z. Fang, C. -Y . Zheng, W.-Q. Tao, Applications of POD -based reduced order model to the rapid prediction of velocity and temperature in data centers, Applied Thermal Engineering, 263 (2025) 125310

2025
[15]

X. Li, Q. Xu, S. Wang, K. Luo, J. Fan, A novel data -driven reduced-order model for the fast prediction of gas -solid heat transfer in fluidized beds, Applied Thermal Engineering, 253 (2024) 123670

2024
[16]

Xiang, F

L. Xiang, F. Wang, Y . Zha, Y . Hu, B. Zhang, X. Yang, R. Hu, X. Luo, Reduced- order-driven recurrent neural network for ultra -fast thermal field simulation in high - heat-flux electronic systems, International Journal of Heat Mass Transfer, 251 (2025) 127351

2025
[17]

F. Wang, Y . Hu, B. Zhang, Y . Zha, X. Luo, Fast Temperature Field Extrapolation Under Non-Periodic Boundary Conditions, Applied Sciences, 15 (2025) 3895

2025
[18]

Chung, Y

S.W. Chung, Y . Choi, P. Roy, T. Moore, T. Roy, T.Y . Lin, D.T. Nguyen, C. Hahn, E.B. Duoss, S.E. Baker, Train small, model big: Scalable physics simulators via reduced order modeling and domain decomposition, Computer Methods in Applied Mechanics and Engineering, 427 (2024) 117041

2024
[19]

R. Fu, D. Xiao, A.G. Buchan, X. Lin, Y . Feng, G. Dong, A parametric non-linear non-intrusive reduce-order model using deep transfer learning, Computer Methods in Applied Mechanics and Engineering, 438 (2025) 117807

2025
[20]

Ali, G.Y

A.H. Ali, G.Y . Mohanad, A. Mohammad, A. Saad Abbas, Transfer Learning: A New Promising Techniques, Mesopotamian Journal of Big Data, 2023 (2023) 29-30

2023
[21]

W. Guo, F. Zhuang, X. Zhang, Y . Tong, J. Dong, A comprehensive survey of federated transfer learning: challenges, methods and applications, Frontiers of Computer Science, 18 (2024) 186356

2024
[22]

Gupta, S

J. Gupta, S. Pathak, G. Kumar, Deep learning (CNN) and transfer learning: a review, in: Journal of Physics: Conference Series, V ol. 2273, IOP Publishing, 2022, pp. 012029

2022
[23]

Zhang, C

S. Zhang, C. Deng, F. Wei, J. Li, X. Luo, J. Ma, Efficient structure optimization of downhole thermal management system via transfer learning surrogate modeling, Journal of Energy Storage, 151 (2026) 120543

2026
[24]

T. Chen, P. Jiang, Decoupling thermal properties in multilayered systems for advanced thermoreflectance experiments, Physical Review Applied, 23 (2025) 044004

2025
[25]

Poopakdee, Z

N. Poopakdee, Z. Abdallah, J.W. Pomeroy, M. Kuball, In situ thermoreflectance characterization of thermal resistance in multilayer electronics packaging, ACS Applied Electronic Materials, 4 (2022) 1558-1566

2022
[26]

Brown, W

D.B. Brown, W. Shen, X. Li, K. Xiao, D.B. Geohegan, S. Kumar, Spatial mapping of thermal boundary conductance at metal –molybdenum diselenide interfaces, ACS Applied Materials Interfaces, 11 (2019) 14418-14426

2019
[27]

B. Xiao, T. Chen, W. Zhang, X. Qian, P. Jiang, Fast and reliable thermal property extraction from FDTR measurements via hybrid particle swarm optimization, International Journal of Heat Mass Transfer, 256 (2026) 128109

2026
[28]

Fan, C.-J

S.-K. Fan, C.-J. Kim, J.-A. Paik, B. Dunn, P.R. Patterson, M.C. Wu, MEMS with thin-film aerogel, in: Technical Digest. MEMS 2001. 14th IEEE International Conference on Micro Electro Mechanical Systems (Cat. No. 01CH37090), IEEE, 2001, pp. 122-125

2001
[29]

Duy Hien, R.A

T. Duy Hien, R.A. Zwijze, J.W. Berenschot, R.J. Wiegerink, G.J.M. Krijnen, M. Elwenspoek, Platinum Patterning by a Modified Lift-Off Technique and Its Application in a Silicon Load Cell, Sensors and Materials, 13 (2001)

2001