Physics-Constrained Denoising Autoencoders for Data-Scarce Wildfire UAV Sensing

Abdelrahman Ramadan; David S. McLagan; Emily Taylor; Lucas Edwards; Melissa Greeff; Sidney Givigi; Xan Giuliani; Zahra Dorbeigi Namaghi

arxiv: 2601.11794 · v1 · submitted 2026-01-16 · 💻 cs.LG · cs.CV· cs.RO

Physics-Constrained Denoising Autoencoders for Data-Scarce Wildfire UAV Sensing

Abdelrahman Ramadan , Zahra Dorbeigi Namaghi , Emily Taylor , Lucas Edwards , Xan Giuliani , David S. McLagan , Sidney Givigi , Melissa Greeff This is my paper

Pith reviewed 2026-05-16 13:02 UTC · model grok-4.3

classification 💻 cs.LG cs.CVcs.RO

keywords denoising autoencoderphysics-informed neural networkUAV sensingwildfire monitoringdata scarcitysensor noise reductionnon-negative constraints

0 comments

The pith

Embedding physical constraints into a denoising autoencoder cleans noisy low-cost UAV wildfire sensor data using only a few thousand samples.

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

The paper shows how to denoise atmospheric measurements from low-cost sensors on UAVs during wildfires without needing the large datasets that standard deep learning requires. It builds non-negativity and temporal smoothing rules straight into the network layers so every output is physically valid by design. On roughly 7900 one-second samples from real prescribed-burn flights, the smallest version of the model improves smoothness by 67 percent and cuts high-frequency noise by 90 percent while producing zero invalid negative readings. Other common architectures generate negative concentrations on the same data.

Core claim

PC²DAE is a physics-informed denoising autoencoder that addresses data scarcity by embedding physical constraints directly into the network architecture. Non-negative concentration estimates are enforced via softplus activations and physically plausible temporal smoothing, ensuring outputs are physically admissible by construction rather than relying on loss function penalties. On 7,894 synchronized 1 Hz samples from UAV flights during prescribed burns, PC²DAE-Lean achieves 67.3% smoothness improvement and 90.7% high-frequency noise reduction with zero physics violations.

What carries the argument

Physics-constrained denoising autoencoder that uses softplus activations to enforce non-negativity and hierarchical decoder heads for separate sensor families together with built-in temporal smoothing.

If this is right

The lean 21k-parameter version trains in under 65 seconds on ordinary hardware and runs on edge devices for real-time UAV use.
The same architecture outperforms larger models on this small dataset because the physics rules act as a strong regularizer against overfitting.
Five standard baselines (LSTM-AE, U-Net, Transformer, CBDAE, DeSpaWN) all generate 15-23% negative outputs on the same data.
The method works with two orders of magnitude less data than typical deep-learning denoising requirements.

Where Pith is reading between the lines

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

The same style of hard architectural constraints could be applied to other scientific sensor-cleaning problems where non-negativity or monotonicity is known a priori.
If the model remains accurate on uncontrolled wildfire events rather than only prescribed burns, it would allow finer-grained plume tracking from small UAV fleets.
Reducing parameter count while adding explicit physical rules may be a general tactic for any domain where labeled scientific data is expensive to collect.

Load-bearing premise

That enforcing non-negativity via softplus activations and temporal smoothing in the architecture fully captures the relevant sensor physics and does not introduce new biases or miss important unmodeled effects in the real wildfire data.

What would settle it

Applying the trained model to a fresh collection of UAV wildfire sensor readings and finding that it produces negative concentration values or shows no reduction in high-frequency noise relative to the raw inputs.

Figures

Figures reproduced from arXiv: 2601.11794 by Abdelrahman Ramadan, David S. McLagan, Emily Taylor, Lucas Edwards, Melissa Greeff, Sidney Givigi, Xan Giuliani, Zahra Dorbeigi Namaghi.

**Figure 2.** Figure 2: Data flow architecture showing multi-sensor integration and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: PC2DAE architecture for physics-constrained sensor denoising. The shared TCN encoder comprises three dilated 1D convolutional blocks with exponentially increasing dilation factors (1, 2, 4) yielding ∼57-sample receptive field matched to sensor response dynamics. The symmetric decoder feeds into family-specific physics-constrained heads (BC: 4 channels, Gas: 9 channels, CO2: 2 channels), each enhanced with … view at source ↗

**Figure 4.** Figure 4: Signal reconstruction comparison for Black Carbon and Gas sensor families. (a) Black Carbon UV channel: Raw input (gray shaded) exhibits [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Multi-dimensional comparison across six evaluation axes: Smooth [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

read the original abstract

Wildfire monitoring requires high-resolution atmospheric measurements, yet low-cost sensors on Unmanned Aerial Vehicles (UAVs) exhibit baseline drift, cross-sensitivity, and response lag that corrupt concentration estimates. Traditional deep learning denoising approaches demand large datasets impractical to obtain from limited UAV flight campaigns. We present PC$^2$DAE, a physics-informed denoising autoencoder that addresses data scarcity by embedding physical constraints directly into the network architecture. Non-negative concentration estimates are enforced via softplus activations and physically plausible temporal smoothing, ensuring outputs are physically admissible by construction rather than relying on loss function penalties. The architecture employs hierarchical decoder heads for Black Carbon, Gas, and CO$_2$ sensor families, with two variants: PC$^2$DAE-Lean (21k parameters) for edge deployment and PC$^2$DAE-Wide (204k parameters) for offline processing. We evaluate on 7,894 synchronized 1 Hz samples collected from UAV flights during prescribed burns in Saskatchewan, Canada (approximately 2.2 hours of flight data), two orders of magnitude below typical deep learning requirements. PC$^2$DAE-Lean achieves 67.3\% smoothness improvement and 90.7\% high-frequency noise reduction with zero physics violations. Five baselines (LSTM-AE, U-Net, Transformer, CBDAE, DeSpaWN) produce 15--23\% negative outputs. The lean variant outperforms wide (+5.6\% smoothness), suggesting reduced capacity with strong inductive bias prevents overfitting in data-scarce regimes. Training completes in under 65 seconds on consumer hardware.

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

PC²DAE bakes non-negativity and smoothing into the architecture for denoising low-cost UAV wildfire sensors and beats baselines on real flight data, but the gains rest on proxy metrics without independent ground truth.

read the letter

The main takeaway is that this work embeds physical constraints like softplus non-negativity and temporal smoothing straight into the decoder of a denoising autoencoder, along with separate heads for Black Carbon, gas, and CO2 sensor families. On 7,894 real 1 Hz samples from prescribed burns, the lean 21k-parameter version delivers 67% better smoothness and 91% less high-frequency noise than the baselines, with zero physics violations and faster training than expected for the data size. The fact that the smaller model beats the wider one is a useful signal that the inductive biases are doing real work in the data-scarce regime they target. That part is cleanly executed and worth noting for anyone building constrained models for edge sensing. The evaluation stays on held-out real flights rather than synthetic data, which keeps the circularity low. The soft spot is the validation. All reported improvements come from internal proxies for smoothness and noise; there are no co-located reference instruments or simulator runs with known true concentrations to measure actual error or bias. It is possible the outputs are admissible yet systematically shifted from the real atmospheric signals. Baseline training details are also thin, so the 15-23% negative outputs they report could partly reflect implementation differences. This is the kind of applied paper that belongs in a reading group focused on physics-informed ML or environmental sensing. It has enough real data and concrete architecture choices to deserve peer review, even if reviewers will press on the lack of external validation and ask for more ablation on the constraint choices.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes PC²DAE, a physics-constrained denoising autoencoder for processing noisy data from low-cost UAV sensors in wildfire monitoring. Physical constraints are embedded architecturally via softplus activations for non-negativity and temporal smoothing, with hierarchical heads for different sensor types. Two model sizes are presented and evaluated on 7,894 real synchronized samples from prescribed burn flights, claiming substantial improvements in smoothness and noise reduction over five baselines while guaranteeing zero physics violations.

Significance. If the results hold, the work is significant for enabling effective denoising in data-scarce regimes typical of UAV-based environmental sensing. The use of strong inductive biases to achieve good performance with only 21k parameters and training in under 65 seconds on consumer hardware is noteworthy, as is the demonstration on real flight data rather than simulations. This could facilitate more reliable wildfire atmospheric measurements with affordable hardware.

major comments (2)

[§4 (Evaluation)] §4 (Evaluation): The central performance claims (67.3% smoothness improvement, 90.7% high-frequency noise reduction) rely exclusively on proxy metrics computed from the real data without any ground-truth reference measurements or co-located high-accuracy sensors. This leaves open the possibility that the outputs are admissible but biased relative to actual concentrations.
[§3 (Architecture)] §3 (Architecture): The assumption that softplus activations and temporal smoothing fully capture the relevant sensor physics (non-negativity and smoothness) may not account for other unmodeled effects such as cross-sensitivity and response lag mentioned in the introduction; additional validation against a physics simulator with known ground truth would strengthen the claim.

minor comments (2)

[Abstract] Abstract: The abstract could more clearly state the number of parameters and training time for the wide variant to allow direct comparison.
[§2 (Related Work)] §2 (Related Work): Ensure all baselines (LSTM-AE, U-Net, etc.) have their training procedures and hyperparameter selection fully detailed to enable reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and insightful comments. We address each major comment below, providing clarifications on our evaluation approach and architectural choices while incorporating revisions to strengthen the manuscript.

read point-by-point responses

Referee: §4 (Evaluation): The central performance claims (67.3% smoothness improvement, 90.7% high-frequency noise reduction) rely exclusively on proxy metrics computed from the real data without any ground-truth reference measurements or co-located high-accuracy sensors. This leaves open the possibility that the outputs are admissible but biased relative to actual concentrations.

Authors: We acknowledge that ground-truth reference measurements are unavailable in our real-world UAV wildfire dataset, as co-located high-accuracy sensors were not deployed during the prescribed burn flights due to logistical and safety constraints in the field. Our proxy metrics (smoothness via total variation and high-frequency noise via FFT analysis) follow standard practice in sensor denoising literature for data-scarce environmental monitoring. To address the concern of potential bias, we have added a new limitations subsection (Section 5.3) explicitly discussing this issue, including quantitative bounds on possible bias derived from sensor specifications, and outlining future work with co-located reference instruments. The zero unphysical outputs and consistent outperformance over baselines remain valid indicators of practical utility. revision: yes
Referee: §3 (Architecture): The assumption that softplus activations and temporal smoothing fully capture the relevant sensor physics (non-negativity and smoothness) may not account for other unmodeled effects such as cross-sensitivity and response lag mentioned in the introduction; additional validation against a physics simulator with known ground truth would strengthen the claim.

Authors: We agree that cross-sensitivity and response lag are relevant effects noted in the introduction. The PC²DAE architecture prioritizes hard enforcement of non-negativity (via softplus) and temporal smoothness as the most critical constraints to eliminate physics violations, which the baselines fail to achieve. Full modeling of cross-sensitivity would require per-sensor calibration datasets not present in our 7,894-sample collection. We have revised Section 3.4 to clarify the scope of embedded physics and added a paragraph in the discussion noting that simulator-based validation with known ground truth is a valuable future direction but would require shifting focus from real data-scarce regimes. No simulator experiments were added, as they fall outside the paper's emphasis on field UAV data. revision: partial

Circularity Check

1 steps flagged

Zero physics violations reported by construction via softplus enforcement

specific steps

self definitional [Abstract]
"Non-negative concentration estimates are enforced via softplus activations and physically plausible temporal smoothing, ensuring outputs are physically admissible by construction rather than relying on loss function penalties. ... PC²DAE-Lean achieves 67.3% smoothness improvement and 90.7% high-frequency noise reduction with zero physics violations."

The 'zero physics violations' is guaranteed by the softplus activations that enforce non-negativity in the architecture; the reported zero is therefore a direct consequence of the design choice and not an empirical result derived from the UAV flight data or training process.

full rationale

The paper embeds non-negativity and temporal smoothing directly into the network architecture (softplus activations and smoothing layers) and then reports 'zero physics violations' as a performance achievement on the 7,894 samples. This specific claim reduces to a definitional property of the chosen architecture rather than an independent outcome of training or evaluation. The smoothness and noise-reduction percentages are computed on held-out real data against external baselines and do not reduce to the inputs by construction, nor are there self-citations, uniqueness theorems, or fitted parameters that force the central results. The overall derivation chain therefore remains largely self-contained, with only this one minor self-definitional element.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard domain assumptions about physical admissibility of concentration readings rather than new fitted parameters or invented entities; training still occurs via standard optimization but the inductive bias is architectural.

axioms (2)

domain assumption Atmospheric concentrations measured by the sensors are non-negative
Invoked via softplus activations to enforce physical admissibility by construction
domain assumption Temporal evolution of concentrations follows physically plausible smooth changes
Enforced directly in the network architecture rather than via loss penalties

pith-pipeline@v0.9.0 · 5626 in / 1382 out tokens · 46619 ms · 2026-05-16T13:02:57.591052+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Non-negative concentration estimates are enforced via softplus activations and physically plausible temporal smoothing, ensuring outputs are physically admissible by construction rather than relying on loss function penalties.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

Auto-Encoders in Deep Learning—A Review with New Perspectives,

S. Chen and W. Guo, “Auto-Encoders in Deep Learning—A Review with New Perspectives,”Mathematics, vol. 11, p. 1777, Jan. 2023. Publisher: Multidisciplinary Digital Publishing Institute. 0 1000 2000 3000 4000 5000 6000 7000 Time (s) 20000 0 20000 40000 60000 80000 UV BC1 (scaled) 7250 7260 7270 7280 7290 7300 7310 Time (s) 20000 0 20000 40000 60000 80000 UV...

work page 2023
[2]

Network of low-cost air quality sensors for monitoring indoor, outdoor, and personal PM2.5 exposure in Seattle during the 2020 wildfire season,

J. He, C.-H. Huang, N. Yuan, E. Austin, E. Seto, and I. Novosselov, “Network of low-cost air quality sensors for monitoring indoor, outdoor, and personal PM2.5 exposure in Seattle during the 2020 wildfire season,” Atmospheric Environment, vol. 285, p. 119244, Sept. 2022

work page 2020
[3]

Wildfire Smoke Air Monitoring Response Technology (WSMART)

ORD. US EPA, “Wildfire Smoke Air Monitoring Response Technology (WSMART).” https://www.epa.gov/air-sensor-toolbox/wildfire-smoke- air-monitoring-response-technology-wsmart, Aug. 2021

work page 2021
[4]

Performance of Vehicle Add-on Mobile Monitoring Sys- tem PM2.5 measurements during wildland fire episodes,

A. S. Bittner, A. L. Holder, A. P. Grieshop, G. S. W. Hagler, and W. Mitchell, “Performance of Vehicle Add-on Mobile Monitoring Sys- tem PM2.5 measurements during wildland fire episodes,”Environmental Science: Atmospheres, vol. 4, pp. 306–320, Mar. 2024. Publisher: RSC

work page 2024
[5]

Challenges and Opportunities in Calibrating Low-Cost Environmental Sensors,

N. V . S. R. Nalakurthi, I. Abimbola, T. Ahmed, I. Anton, K. Riaz, Q. Ibrahim, A. Banerjee, A. Tiwari, and S. Gharbia, “Challenges and Opportunities in Calibrating Low-Cost Environmental Sensors,”Sensors (Basel, Switzerland), vol. 24, p. 3650, June 2024

work page 2024
[6]

Field evaluation of low-cost electrochemical air quality gas sensors under extreme temperature and relative humidity conditions,

R. Papaconstantinou, M. Demosthenous, S. Bezantakos, N. Hadji- georgiou, M. Costi, M. Stylianou, E. Symeou, C. Savvides, and G. Biskos, “Field evaluation of low-cost electrochemical air quality gas sensors under extreme temperature and relative humidity conditions,” Atmospheric Measurement Techniques, vol. 16, pp. 3313–3329, June 2023

work page 2023
[7]

Targeted Adversarial Denoising Autoencoders (TADA) for Neural Time Series Filtration

“Targeted Adversarial Denoising Autoencoders (TADA) for Neural Time Series Filtration.” https://arxiv.org/html/2501.04967

work page arXiv
[8]

Wasserstein Auto-Encoders,

I. Tolstikhin, O. Bousquet, S. Gelly, and B. Schoelkopf, “Wasserstein Auto-Encoders,” inInternational Conference on Learning Representa- tions, Feb. 2018

work page 2018
[9]

Physics-Informed Masked Autoencoder for active sparse imaging,

L. McEvoy, D. Tafone, Y . M. Sua, and Y . Huang, “Physics-Informed Masked Autoencoder for active sparse imaging,”Scientific Reports, vol. 14, p. 20078, Aug. 2024

work page 2024
[10]

Physics-informed neural networks for parameter learning of wildfire spreading,

K. V ogiatzoglou, C. Papadimitriou, V . Bontozoglou, and K. Ampoun- tolas, “Physics-informed neural networks for parameter learning of wildfire spreading,”Computer Methods in Applied Mechanics and Engineering, vol. 434, p. 117545, Feb. 2025

work page 2025
[11]

An advanced physics-informed neural network-based frame- work for nonlinear and complex topology optimization,

H. Jeong, C. Batuwatta-Gamage, J. Bai, C. Rathnayaka, Y . Zhou, and Y . Gu, “An advanced physics-informed neural network-based frame- work for nonlinear and complex topology optimization,”Engineering Structures, vol. 322, p. 119194, Jan. 2025

work page 2025
[12]

Mechanics-informed autoencoder enables automated detection and localization of unforeseen structural damage,

X. Li, H. Bolandi, M. Masmoudi, T. Salem, A. Jha, N. Lajnef, and V . N. Boddeti, “Mechanics-informed autoencoder enables automated detection and localization of unforeseen structural damage,”Nature Communications, vol. 15, p. 9229, Oct. 2024

work page 2024
[13]

Physics and semantic informed multi-sensor calibration via optimization theory and self-supervised learning,

S. Y . Hayoun, M. Halachmi, D. Serebro, K. Twizer, E. Medezinski, L. Korkidi, M. Cohen, and I. Orr, “Physics and semantic informed multi-sensor calibration via optimization theory and self-supervised learning,”Scientific Reports, vol. 14, p. 2541, Jan. 2024

work page 2024
[14]

Physics-Informed deep Autoencoder for fault detection in New-Design systems,

C. Lai, P. Baraldi, and E. Zio, “Physics-Informed deep Autoencoder for fault detection in New-Design systems,”Mechanical Systems and Signal Processing, vol. 215, p. 111420, June 2024

work page 2024
[15]

Correcting for background currents in four electrode toxic gas sensors,

Alphasense Ltd., “Correcting for background currents in four electrode toxic gas sensors,” Application Note AAN 803, Alphasense Ltd., Great Notley, Essex, UK, Mar. 2019

work page 2019
[16]

Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series,

G. Michau, G. Frusque, and O. Fink, “Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series,” Proceedings of the National Academy of Sciences, vol. 119, no. 8, p. e2106598119, 2022

work page 2022
[17]

Contrastive blind denoising autoencoder for real time denoising of industrial IoT sensor data,

S. Langarica and F. N´u˜nez, “Contrastive blind denoising autoencoder for real time denoising of industrial IoT sensor data,” vol. 120, p. 105838

work page

[1] [1]

Auto-Encoders in Deep Learning—A Review with New Perspectives,

S. Chen and W. Guo, “Auto-Encoders in Deep Learning—A Review with New Perspectives,”Mathematics, vol. 11, p. 1777, Jan. 2023. Publisher: Multidisciplinary Digital Publishing Institute. 0 1000 2000 3000 4000 5000 6000 7000 Time (s) 20000 0 20000 40000 60000 80000 UV BC1 (scaled) 7250 7260 7270 7280 7290 7300 7310 Time (s) 20000 0 20000 40000 60000 80000 UV...

work page 2023

[2] [2]

Network of low-cost air quality sensors for monitoring indoor, outdoor, and personal PM2.5 exposure in Seattle during the 2020 wildfire season,

J. He, C.-H. Huang, N. Yuan, E. Austin, E. Seto, and I. Novosselov, “Network of low-cost air quality sensors for monitoring indoor, outdoor, and personal PM2.5 exposure in Seattle during the 2020 wildfire season,” Atmospheric Environment, vol. 285, p. 119244, Sept. 2022

work page 2020

[3] [3]

Wildfire Smoke Air Monitoring Response Technology (WSMART)

ORD. US EPA, “Wildfire Smoke Air Monitoring Response Technology (WSMART).” https://www.epa.gov/air-sensor-toolbox/wildfire-smoke- air-monitoring-response-technology-wsmart, Aug. 2021

work page 2021

[4] [4]

Performance of Vehicle Add-on Mobile Monitoring Sys- tem PM2.5 measurements during wildland fire episodes,

A. S. Bittner, A. L. Holder, A. P. Grieshop, G. S. W. Hagler, and W. Mitchell, “Performance of Vehicle Add-on Mobile Monitoring Sys- tem PM2.5 measurements during wildland fire episodes,”Environmental Science: Atmospheres, vol. 4, pp. 306–320, Mar. 2024. Publisher: RSC

work page 2024

[5] [5]

Challenges and Opportunities in Calibrating Low-Cost Environmental Sensors,

N. V . S. R. Nalakurthi, I. Abimbola, T. Ahmed, I. Anton, K. Riaz, Q. Ibrahim, A. Banerjee, A. Tiwari, and S. Gharbia, “Challenges and Opportunities in Calibrating Low-Cost Environmental Sensors,”Sensors (Basel, Switzerland), vol. 24, p. 3650, June 2024

work page 2024

[6] [6]

Field evaluation of low-cost electrochemical air quality gas sensors under extreme temperature and relative humidity conditions,

R. Papaconstantinou, M. Demosthenous, S. Bezantakos, N. Hadji- georgiou, M. Costi, M. Stylianou, E. Symeou, C. Savvides, and G. Biskos, “Field evaluation of low-cost electrochemical air quality gas sensors under extreme temperature and relative humidity conditions,” Atmospheric Measurement Techniques, vol. 16, pp. 3313–3329, June 2023

work page 2023

[7] [7]

Targeted Adversarial Denoising Autoencoders (TADA) for Neural Time Series Filtration

“Targeted Adversarial Denoising Autoencoders (TADA) for Neural Time Series Filtration.” https://arxiv.org/html/2501.04967

work page arXiv

[8] [8]

Wasserstein Auto-Encoders,

I. Tolstikhin, O. Bousquet, S. Gelly, and B. Schoelkopf, “Wasserstein Auto-Encoders,” inInternational Conference on Learning Representa- tions, Feb. 2018

work page 2018

[9] [9]

Physics-Informed Masked Autoencoder for active sparse imaging,

L. McEvoy, D. Tafone, Y . M. Sua, and Y . Huang, “Physics-Informed Masked Autoencoder for active sparse imaging,”Scientific Reports, vol. 14, p. 20078, Aug. 2024

work page 2024

[10] [10]

Physics-informed neural networks for parameter learning of wildfire spreading,

K. V ogiatzoglou, C. Papadimitriou, V . Bontozoglou, and K. Ampoun- tolas, “Physics-informed neural networks for parameter learning of wildfire spreading,”Computer Methods in Applied Mechanics and Engineering, vol. 434, p. 117545, Feb. 2025

work page 2025

[11] [11]

An advanced physics-informed neural network-based frame- work for nonlinear and complex topology optimization,

H. Jeong, C. Batuwatta-Gamage, J. Bai, C. Rathnayaka, Y . Zhou, and Y . Gu, “An advanced physics-informed neural network-based frame- work for nonlinear and complex topology optimization,”Engineering Structures, vol. 322, p. 119194, Jan. 2025

work page 2025

[12] [12]

Mechanics-informed autoencoder enables automated detection and localization of unforeseen structural damage,

X. Li, H. Bolandi, M. Masmoudi, T. Salem, A. Jha, N. Lajnef, and V . N. Boddeti, “Mechanics-informed autoencoder enables automated detection and localization of unforeseen structural damage,”Nature Communications, vol. 15, p. 9229, Oct. 2024

work page 2024

[13] [13]

Physics and semantic informed multi-sensor calibration via optimization theory and self-supervised learning,

S. Y . Hayoun, M. Halachmi, D. Serebro, K. Twizer, E. Medezinski, L. Korkidi, M. Cohen, and I. Orr, “Physics and semantic informed multi-sensor calibration via optimization theory and self-supervised learning,”Scientific Reports, vol. 14, p. 2541, Jan. 2024

work page 2024

[14] [14]

Physics-Informed deep Autoencoder for fault detection in New-Design systems,

C. Lai, P. Baraldi, and E. Zio, “Physics-Informed deep Autoencoder for fault detection in New-Design systems,”Mechanical Systems and Signal Processing, vol. 215, p. 111420, June 2024

work page 2024

[15] [15]

Correcting for background currents in four electrode toxic gas sensors,

Alphasense Ltd., “Correcting for background currents in four electrode toxic gas sensors,” Application Note AAN 803, Alphasense Ltd., Great Notley, Essex, UK, Mar. 2019

work page 2019

[16] [16]

Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series,

G. Michau, G. Frusque, and O. Fink, “Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series,” Proceedings of the National Academy of Sciences, vol. 119, no. 8, p. e2106598119, 2022

work page 2022

[17] [17]

Contrastive blind denoising autoencoder for real time denoising of industrial IoT sensor data,

S. Langarica and F. N´u˜nez, “Contrastive blind denoising autoencoder for real time denoising of industrial IoT sensor data,” vol. 120, p. 105838

work page