Periodic Reporting for period 1 - GEOLEARN (Real-time hydrogen-storage monitoring via energy-efficient deep learning)
Reporting period: 2024-01-29 to 2026-01-28
The adoption of renewable hydrogen as green fuel and energy storage is key to
achieve the European Green Deal. However, its large-scale storage is still
facing significant challenges. Deep learning (DL) measurement inversion is a
state-of-the-art approach used for underground storage-site monitoring and
detection. However: 1) It requires a huge amount of training data. 2) DL
training is incredibly expensive. 3) While multiscale data are often available,
there are no DL techniques for multiscale electromagnetic measurement
inversion.
The goal of GEOLEARN is to guide hydrogen storage technologies by inverting
subsurface multiscale electromagnetic measurements in real time using
energy-efficient DL methods. For this purpose, GEOLEARN aims to leverage
mixed-precision (MP) computations to maximise energy- and cost-efficiency, and
ensure scalability. The objectives of GEOLEARN target the above challenges and
read as follows: 1) Develop MP finite element methods (FEM) that can rapidly
generate large training data. 2) Design MP DL algorithms that can efficiently
process huge databases during training and invert measurements in real time. 3)
Apply the new techniques to invert multiscale geophysical electromagnetic
measurements and guide hydrogen storage.
achieve the European Green Deal. However, its large-scale storage is still
facing significant challenges. Deep learning (DL) measurement inversion is a
state-of-the-art approach used for underground storage-site monitoring and
detection. However: 1) It requires a huge amount of training data. 2) DL
training is incredibly expensive. 3) While multiscale data are often available,
there are no DL techniques for multiscale electromagnetic measurement
inversion.
The goal of GEOLEARN is to guide hydrogen storage technologies by inverting
subsurface multiscale electromagnetic measurements in real time using
energy-efficient DL methods. For this purpose, GEOLEARN aims to leverage
mixed-precision (MP) computations to maximise energy- and cost-efficiency, and
ensure scalability. The objectives of GEOLEARN target the above challenges and
read as follows: 1) Develop MP finite element methods (FEM) that can rapidly
generate large training data. 2) Design MP DL algorithms that can efficiently
process huge databases during training and invert measurements in real time. 3)
Apply the new techniques to invert multiscale geophysical electromagnetic
measurements and guide hydrogen storage.
Following its ambitions, the GEOLEARN project has developed new mathematical
theories and computational techniques which pave the way towards real-time
subsurface hydrogen storage monitoring. The project has so far focused on the
development of mixed-precision (MP) finite element methods (FEMs) for accelerating
and increasing the energy efficiency of computations. This topic
had never been investigated before.
GEOLEARN started by developing the underlying mathematical theory of MP FEM
computations. The resulting rounding error analysis provides a theoretical
framework that can guide implementations and future developments. Informed by
the new theory, the project designed new MP FEM algorithms and software that
exploit the MP hardware accelerators of artificial intelligence (AI) chips,
resulting in significant (up to 60x) computational speedups.
A challenge in scientific computing is the exploitation of AI chips for non-AI
computational tasks. The project results demonstrate how such hardware can be
employed to improve performance even on well-established computational
frameworks such as the FEM. At the same time, GEOLEARN has provided new
evidence that such improvements are only possible provided that the
implementation is backed by a rigorous theoretical analysis. Overall, these
results provide a motivating success story for further adoption in other
scientific computing methodologies and will likely open further avenues of
investigation.
Through research visits, organisation of minisymposia, and participation in
international conferences, the GEOLEARN team has successfully established
research collaborations between the host institution and other international
researchers. These new contacts have created a vibrant research network
which will facilitate the subsequent GEOLEARN objectives.
theories and computational techniques which pave the way towards real-time
subsurface hydrogen storage monitoring. The project has so far focused on the
development of mixed-precision (MP) finite element methods (FEMs) for accelerating
and increasing the energy efficiency of computations. This topic
had never been investigated before.
GEOLEARN started by developing the underlying mathematical theory of MP FEM
computations. The resulting rounding error analysis provides a theoretical
framework that can guide implementations and future developments. Informed by
the new theory, the project designed new MP FEM algorithms and software that
exploit the MP hardware accelerators of artificial intelligence (AI) chips,
resulting in significant (up to 60x) computational speedups.
A challenge in scientific computing is the exploitation of AI chips for non-AI
computational tasks. The project results demonstrate how such hardware can be
employed to improve performance even on well-established computational
frameworks such as the FEM. At the same time, GEOLEARN has provided new
evidence that such improvements are only possible provided that the
implementation is backed by a rigorous theoretical analysis. Overall, these
results provide a motivating success story for further adoption in other
scientific computing methodologies and will likely open further avenues of
investigation.
Through research visits, organisation of minisymposia, and participation in
international conferences, the GEOLEARN team has successfully established
research collaborations between the host institution and other international
researchers. These new contacts have created a vibrant research network
which will facilitate the subsequent GEOLEARN objectives.
Until its early termination after seven months from the project start date, the
GEOLEARN project has brought its first objective to completion and successfully
developed new mixed-precision (MP) finite element methods (FEM). The
application of MP techniques to the FEM had never been investigated beforehand.
In particular, the GEOLEARN project has:
a) Derived the first rounding error analysis of FEM kernels and assembly. This
theory, submitted for publication, also accounts for MP implementations and MP
hardware accelerators, thus providing theoretical guidance for designing MP FEM
algorithms.
b) Designed new MP FEM implementation strategies and MP hardware-accelerated
algorithms which are provenly accurate as well as up to 60 times faster than
standard (double-precision) implementations. This work has been submitted for
publication and the resulting software is open source and freely accessible
online.
c) Identified new challenges and research opportunities in MP FEM algorithmic
design. Indeed, further research is still required to establish
production-ready MP FEMs with a mature enough technology to be incorporated
into open-source and commercial FEM software. Nevertheless, the project new
MP FEM algorithms are an important stepping stone towards this goal.
d) Created new inter-disciplinary and inter-university collaborations and
opened new avenues of investigation on MP methods.
e) Disseminated project results and ideas by organising minisymposia, giving talks
at conferences, and performing research visits.
The importance of this work lies in the fact that FEMs are ubiquitously used in
scientific computing and the widespread adoption of the new methods by the
scientific community can bring wide scientific impact.
GEOLEARN project has brought its first objective to completion and successfully
developed new mixed-precision (MP) finite element methods (FEM). The
application of MP techniques to the FEM had never been investigated beforehand.
In particular, the GEOLEARN project has:
a) Derived the first rounding error analysis of FEM kernels and assembly. This
theory, submitted for publication, also accounts for MP implementations and MP
hardware accelerators, thus providing theoretical guidance for designing MP FEM
algorithms.
b) Designed new MP FEM implementation strategies and MP hardware-accelerated
algorithms which are provenly accurate as well as up to 60 times faster than
standard (double-precision) implementations. This work has been submitted for
publication and the resulting software is open source and freely accessible
online.
c) Identified new challenges and research opportunities in MP FEM algorithmic
design. Indeed, further research is still required to establish
production-ready MP FEMs with a mature enough technology to be incorporated
into open-source and commercial FEM software. Nevertheless, the project new
MP FEM algorithms are an important stepping stone towards this goal.
d) Created new inter-disciplinary and inter-university collaborations and
opened new avenues of investigation on MP methods.
e) Disseminated project results and ideas by organising minisymposia, giving talks
at conferences, and performing research visits.
The importance of this work lies in the fact that FEMs are ubiquitously used in
scientific computing and the widespread adoption of the new methods by the
scientific community can bring wide scientific impact.