Previous Efforts

Data Science at Aquafil Group

About Aquafil

The company operates in the chemical-textile manufacturing industry, with a strong focus on sustainability. Its flagship product, ECONYL®, is a synthetic fiber made from regenerated nylon sourced from end-of-life carpets, textile scraps, and fishing nets. The regeneration process may be infinitely repeated to retrieve the basic raw material. This reflects the company’s commitment to driving positive change in environmental sustainability.

The experience

Aquafil has been my first employer after the academic start of career. The value of this experience has been the first person investment on real-world problem solving. Due to the nature of real data, I could test the limits of theory and distill solutions backed by both theoretical (algorithmic, mathematical physics-based) insights and field-specific operational knowledge. My work was informed by both previously acquired academic know-that and the constant exchange with expert field operators. Their know-how was the bridge between the model and its performance and robustness in a real operational scenario.

My contribution

My contribution to the company involves leveraging proprietary data and scientific knowledge to develop explanatory and predictive models, and ready-to-use software tools. The value I bring is focused on automating and harmonizing quality control processes, analyzing chemical-physical processes.

My role in the scope of Aquafil’s mission is to help in digitalization process. In my case, this means

• to propose and implement data-driven explanatory and predictive models for the automation, harmonization of quality control processes
• to analyze and obtain information from proprietary data sets. These data come from the observation of physical-chemical processes, production machines and quality control tests
• to provide ready-to-use software solutions to assist sector-specific operations

The experience has been stimulating and instructive. This is due to both the mentioned real-world manufacturing processes exposure and the possibility I had to research and to work problems out in full autonomy, thanks to the trust and support of my co-workers and hierarchical supervisors.

Please note that due to NDA constraints, I can not share a high-resolution level of detail on the activities I have been deployed on.

Doctorate at IMT Atlantique

Scientific scope

Focused on ocean data processing, my Ph.D. Thesis aimed to answer the question:

“How can we exploit the complementary information conveyed by diverse, multi-sensor and spatio-temporally heterogeneous data to improve the characterization of the sea surface state?”

This problem was framed as a geophysical inverse problem. That is, to retrieve a bigger picture from its partial observations. In this case, the bigger picture is the dynamical behavior of sea-surface wind speed and the partial observations are the output of the measurement methods that can be use to gauge the phenomenon. Wind speed at the sea surface is an important parameter for a vast range of anthropic activities. The spatial extent of the sea surface and the strong spatio-temporal variability of wind speed makes each individual measurement method undercomplete and uncapable of capturing all of wind speed complexity.

Methodological framework

The methodological approach involved an end-to-end trainable framework, the 4DVarNet. This framework bridges traditional physics-based geophysical inversion schemes and deep learning modelling. This allows us to exploit the geoscientific theoretical foundations, the flexibility and representation power of deep learning and the availability of large data sets. In the case of my work, the further challenge is to leverage multimodal datasets. The joint model-based and data-driven approach allows to use data-driven methods to learn complex, high-dimensional dynamical systems. For reference, see

• the core 4DVarNet repository
• the starter version
• and the original paper

Note. I did not contribute to the development of 4DVarNet. The original authors are listed in the repository and in the publication mentioned above.

Acknowledgements

My Thesis was funded by Agence National de la Recherche (ANR) and co-funded by Naval Group. My thesis work, framed in the context of the project “AI Chair OceaniX”, was supervised by Prof. Ronan Fablet as main supervisor; and by Prof. Nicolas Farrugia and Prof. Dorian Cazau as co-supervisors.

Relevant experiences

IEEE OCEANS 2023-Limerick Conference

Attended in June 2023 in Limerick, Ireland.

Company visit at Naval Group, 83200 Toulon, France

A brief stay at Naval Group, from May to June 2023

Teaching activities support

The Ph.D. students had the possibility to support the course teachers in the didactical activities.
Spring 2021, 2022 and 2023.

Contributions

We contributed with two original papers and one conference proceeding. In the following, a brief description of the main work axes.

Multimodal Learning-based Reconstruction of High-resolution Spatial Wind Speed Fields

Published in January 2025, the work has been done between 2022 and 2023.
Original article in Environmental Data Science.
Arxiv preprint, posted in December 2023.

In this work, we used the 4DVarNet framework to reconstruct high-resolution sea-surface wind speed fields. Our goal was to leverage wind speed data across multiple spatiotemporal scales to maximize the contribution of each input source. Using simulated data, we integrated diverse pseudo-observations—including satellite images, in situ measurements, and reanalysis products—to capture wind speed variability more effectively. Extensive numerical experiments show that this multimodal approach outperforms traditional deep learning inversion methods by successfully combining complementary information from various inputs. We also prove that the scheme may benefice of further capacity by processing separately the automatically learned features of the heterogeneous partial observations.

Learning-Based Temporal Estimation of In-Situ Wind Speed From Underwater Passive Acoustics

Published in August 2023, the work has been done in 2022.
Original article in IEEE Journal of Oceanic Engineering.

In this work, we aimed to reconstruct the sea-surface wind speed using underwater passive acoustics. This problem is, unlike the one presented above, local in space. We build on previous work that leveraged machine learning techniques to invert underwater acoustics to estimate wind speed. The novelty proposed by our work consisted in the reconstruction of the wind speed time-related dynamical patterns. The 4DVarNet scheme represents a suitable choice given its physical foundations. Our experiments on real data demonstrate that the our proposed approach outperforms existing data-driven techniques, achieving up to a 16% reduction in RMSE on wind speed reconstruction. The study emphasizes the role of temporal dynamics in underwater acoustic data and suggests that incorporating multimodal information can further enhance robustness, particularly in cases of missing acoustic data.

Research Fellowship at CCNL

Scientific scope

My research at CCNL was focused on the development of a biologically plausible learning algorithm for a class of deep learning models, Deep Belief Networks (DBN). The interest in this particular architecture stems from its bottom-up data processing and top-down generative capabilities. This scheme makes it particularly appealing to use DBN because of their conceptual similitude to human brains visual neural networks.

The question that my work aimed to answer is

Can we use a DBN to emulate the development of the numerosity perception skill in humans? Do DBN follow a similar numerosity discrimination pattern?

Achivement

The outcomes of our research have been published in the following original paper.

A Developmental Approach for Training Deep Belief Networks

Published in December 2022, the work has been done between in 2020.
Original article in Cognitive Computation.

In this work, we proposed an alternative training algorithm for DBNs, the iterative-DBN (iDBN). Unlike classical DBNs, which are trained in a greedy layer-wise way, our proposed alternative better emulates the bottom-up visual stimulus processing mechanism of human brain visual cortex. A first part of the study compares our algorithm with the classical layer-wise DBN. This part serves the purpose of showing the overall performance alignment and the equivalency of the learned features. Next, we assess the numerosity discrimantion skill, that is how the model can recognize the number of objects in a visual stimulus. In our proposed model the numerosity discrimination pattern aligns well with the one exhibited by human learners, and is robust against catastrophic forgetting scenarios.