Skip to content

Commit 7ac098b

Browse files
naveenguptaiitrnaveenguptaiitr
authored
Update index.html
1 parent b151d2f commit 7ac098b

File tree

1 file changed

+156
-154
lines changed

1 file changed

+156
-154
lines changed

‎projects/GFI-framework/index.html‎

Lines changed: 156 additions & 154 deletions
Original file line numberDiff line numberDiff line change
@@ -23,168 +23,170 @@
2323
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
2424
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans+Display:ital,wght@0,400;0,700;1,400;1,700&display=swap" rel="stylesheet">
2525
</head>
26+
2627
<body>
28+
<main>
2729

28-
<!-- Navigation -->
29-
<nav class="navbar" role="navigation">
30-
<div class="navbar-menu is-active">
31-
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
32-
<a class="navbar-item" href="https://kgml-lab.github.io/">
33-
<span class="icon">
34-
<i class="fas fa-home"></i>
35-
</span>
36-
</a>
37-
</div>
38-
</div>
39-
</nav>
40-
41-
<!-- Title + Authors -->
42-
<section class="hero">
43-
<div class="hero-body">
44-
<div class="container has-text-centered">
45-
<h1 class="title is-1">A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations</h1>
46-
<hr style="width: 60%; margin: 1.5rem auto;">
47-
<h2 class="subtitle is-4 has-text-weight-bold">ICLR 2025</h2>
48-
<div class="is-size-5 publication-authors">
49-
<span class="author-block"><a href="https://www.linkedin.com/in/naveengupta1729/">Naveen Gupta</a><sup>1*</sup>,</span>
50-
<span class="author-block"><a href="https://sawhney-medha.github.io/">Medha Sawhney</a><sup>1*</sup>,</span>
51-
<span class="author-block"><a href="https://arkadaw9.github.io/">Arka Daw</a><sup>2*</sup>,</span>
52-
<span class="author-block"><a href="https://sites.google.com/site/youzuolin044/home">Youzuo Lin</a><sup>3</sup>,</span>
53-
<span class="author-block"><a href="https://anujkarpatne.github.io/">Anuj Karpatne</a><sup>1</sup></span>
54-
</div>
55-
<div class="is-size-5 publication-authors">
56-
<span class="author-block"><sup>1</sup>Virginia Tech,</span>
57-
<span class="author-block"><sup>2</sup>Oak Ridge National Lab,</span>
58-
<span class="author-block"><sup>3</sup>UNC at Chapel Hill</span>
59-
</div>
60-
61-
<div class="buttons is-centered mt-4">
62-
<a href="https://arxiv.org/abs/2410.11247" class="button is-dark is-rounded">
63-
<span class="icon"><i class="ai ai-arxiv"></i></span>
64-
<span>arXiv</span>
65-
</a>
66-
<a href="https://openreview.net/forum?id=yIlyHJdYV3" class="button is-dark is-rounded">
67-
<span class="icon"><i class="fas fa-star"></i></span>
68-
<span>OpenReview</span>
69-
</a>
70-
<a href="https://github.com/KGML-lab/Generalized-Forward-Inverse-Framework-for-DL4SI" class="button is-dark is-rounded">
71-
<span class="icon"><i class="fab fa-github"></i></span>
72-
<span>Code</span>
73-
</a>
74-
<a href="https://iclr.cc/virtual/2025/poster/27736" class="button is-dark is-rounded">
75-
<span class="icon"><i class="fas fa-file-pdf"></i></span>
76-
<span>PDF</span>
30+
<!-- Navigation -->
31+
<nav class="navbar" role="navigation">
32+
<div class="navbar-menu is-active">
33+
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
34+
<a class="navbar-item" href="https://kgml-lab.github.io/">
35+
<span class="icon">
36+
<i class="fas fa-home"></i>
37+
</span>
7738
</a>
7839
</div>
7940
</div>
80-
</div>
81-
</section>
41+
</nav>
8242

83-
<section class="section">
84-
<div class="gfi_framework">
85-
<figure>
86-
<img src="./static/images/GFI_Framework.png" alt="GFI Framework" loading="lazy" width=25%>
87-
<figcaption> Figure 1: A unified framework for solving forward and inverse problems in subsurface imaging. </figcaption>
88-
</figure>
89-
</div>
90-
<hr style="width: 60%; margin: 2rem auto;">
91-
</section>
92-
93-
<!-- Abstract -->
94-
<section class="section">
95-
<div class="container is-max-desktop has-text-justified">
96-
<h2 class="title is-3"> Abstract </h2>
97-
<p>
98-
In subsurface imaging, learning the mapping from velocity maps to seismic waveforms (forward problem) and waveforms to velocity (inverse problem) is important for several applications. While traditional techniques for solving forward and inverse problems are computationally prohibitive, there is a growing interest in leveraging recent advances in deep learning to learn the mapping between velocity maps and seismic waveform images directly from data.
99-
Despite the variety of architectures explored in previous works, several open questions remain unanswered such as the effect of latent space sizes, the importance of manifold learning, the complexity of translation models, and the value of jointly solving forward and inverse problems.
100-
We propose a unified framework to systematically characterize prior research in this area termed the Generalized Forward-Inverse (GFI) framework, building on the assumption of manifolds and latent space translations.
101-
We show that GFI encompasses previous works in deep learning for subsurface imaging, which can be viewed as specific instantiations of GFI.
102-
We also propose two new model architectures within the framework of GFI: Latent U-Net and Invertible X-Net, leveraging the power of U-Nets for domain translation and the ability of IU-Nets to simultaneously learn forward and inverse translations, respectively.
103-
We show that our proposed models achieve state-of-the-art performance for forward and inverse problems on a wide range of synthetic datasets and also investigate their zero-shot effectiveness on two real-world-like datasets.
104-
</p>
105-
</div>
106-
<hr style="width: 60%; margin: 2rem auto;">
107-
</section>
108-
109-
<section class="section">
110-
<div class="container is-max-desktop has-text-justified">
111-
<h2 class="title is-3">Method Overview</h2>
112-
<p>
113-
We propose Generalized Forward-Inverse (GFI) framework based on two assumptions. First, according to the manifold assumption, we assume that the velocity maps v ∈ v&#119985; and seismic
114-
waveforms p ∈ p&#119985; can be projected to their corresponding latent space representations, v&#771; and p&#771;, respectively, which can be mapped back to their reconstructions in the original space, v&#770; and p&#770;.
115-
Note that the sizes of the latent spaces can be smaller or larger than the original spaces. Further, the size of v&#771; may not match with the size of p&#771;. Second, according to the latent space
116-
translation assumption, we assume that the problem of learning forward and inverse mappings in the original spaces of velocity and waveforms can be reformulated as learning translations in their
117-
latent spaces.
118-
</p>
119-
<ol>
120-
<li>
121-
<b>Latent U-Net Architecture: </b> We propose a novel architecture to solve forward and inverse problems using two latent space translation models implemented using U-Nets, termed Latent U-Net. Latent
122-
U-Net uses ConvNet backbones for both encoder-decoder pairs: <code>E</code><sub>v</sub>, <code>D</code><sub>v</sub> and <code>E</code><sub>p</sub>, <code>D</code><sub>p</sub>, to project
123-
v and p to lower-dimensional representations. We also constrain the sizes of the latent spaces of
124-
v&#771; and p&#771;to be identical, i.e., dim(v&#771;) = dim(p&#771;), so that we can train two separate U-Net models to implement the latent space mappings L<sub>v&#771; &rarr; p&#771;</sub> and L<sub>p&#771; &rarr; v&#771;</sub>.
125-
126-
<div class="latent_unet">
127-
<figure>
128-
<img src="./static/images/LatentU-Net.png" alt="Latent U-Net architecture" loading="lazy" width=45%>
129-
<figcaption> Latent U-Net architecture </figcaption>
130-
</figure>
43+
<!-- Title + Authors -->
44+
<section class="hero">
45+
<div class="hero-body">
46+
<div class="container has-text-centered">
47+
<h1 class="title is-1">A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations</h1>
48+
<hr style="width: 60%; margin: 1.5rem auto;">
49+
<h2 class="subtitle is-4 has-text-weight-bold">ICLR 2025</h2>
50+
<div class="is-size-5 publication-authors">
51+
<span class="author-block"><a href="https://www.linkedin.com/in/naveengupta1729/">Naveen Gupta</a><sup>1*</sup>,</span>
52+
<span class="author-block"><a href="https://sawhney-medha.github.io/">Medha Sawhney</a><sup>1*</sup>,</span>
53+
<span class="author-block"><a href="https://arkadaw9.github.io/">Arka Daw</a><sup>2*</sup>,</span>
54+
<span class="author-block"><a href="https://sites.google.com/site/youzuolin044/home">Youzuo Lin</a><sup>3</sup>,</span>
55+
<span class="author-block"><a href="https://anujkarpatne.github.io/">Anuj Karpatne</a><sup>1</sup></span>
13156
</div>
132-
133-
</li>
134-
<li>
135-
<b>Invertible X-Net Architecture: </b> We propose another novel architecture within the GFI framework termed Invertible X-Net to answer
136-
the question: “can we learn a single latent space translation model that can simultaneously solve
137-
both forward and inverse problems?” We employ invertible U-Net in the latent spaces of velocity and waveforms, which can be constrained to be of the same size (just like Latent-UNets), i.e.,
138-
dim(v&#771;) = dim(p&#771;).
139-
<div class="inv_xnet">
140-
<figure>
141-
<img src="./static/images/InvertibleX-Net.png" alt="Invertible X-Net architecture" loading="lazy" width=45%>
142-
<figcaption> Invertible X-Net architecture </figcaption>
143-
</figure>
57+
<div class="is-size-5 publication-authors">
58+
<span class="author-block"><sup>1</sup>Virginia Tech,</span>
59+
<span class="author-block"><sup>2</sup>Oak Ridge National Lab,</span>
60+
<span class="author-block"><sup>3</sup>UNC at Chapel Hill</span>
14461
</div>
145-
</li>
146-
</ol>
147-
</div>
148-
<hr style="width: 60%; margin: 2rem auto;">
149-
</section>
150-
151-
<!-- Results Table Placeholder -->
152-
<section class="section">
153-
<div class="container is-max-desktop">
154-
<h2 class="title is-3">Experiments</h2>
155-
</div>
156-
</section>
157-
158-
<!-- BibTeX -->
159-
<section class="section" id="bibtex">
160-
<div class="container is-max-desktop content">
161-
<h2 class="title">BibTeX</h2>
162-
<pre><code>@inproceedings{
163-
gupta2025a,
164-
title={A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations},
165-
author={Naveen Gupta and Medha Sawhney and Arka Daw and Youzuo Lin and Anuj Karpatne},
166-
booktitle={The Thirteenth International Conference on Learning Representations},
167-
year={2025},
168-
url={https://openreview.net/forum?id=yIlyHJdYV3}
169-
}</code></pre>
170-
</div>
171-
</section>
172-
173-
<!-- Footer -->
174-
<section class="section" id="acknowledgement">
175-
<div class="container is-max-desktop content">
176-
<h2 class="title">Acknowledgement</h2>
177-
<p>
178-
This work was supported in part by NSF awards IIS-2239328 and IIS-2107332. We are grateful
179-
to the Advanced Research Computing (ARC) Center at Virginia Tech for providing access to GPU
180-
compute resources for this project. This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government
181-
retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the
182-
published form of this manuscript, or allow others to do so, for US government purposes. DOE will
183-
provide public access to these results of federally sponsored research in accordance with the DOE
184-
Public Access Plan ( https://www.energy.gov/doe-public-access-plan).
185-
</p>
186-
</div>
187-
</footer>
62+
63+
<div class="buttons is-centered mt-4">
64+
<a href="https://arxiv.org/abs/2410.11247" class="button is-dark is-rounded">
65+
<span class="icon"><i class="ai ai-arxiv"></i></span>
66+
<span>arXiv</span>
67+
</a>
68+
<a href="https://openreview.net/forum?id=yIlyHJdYV3" class="button is-dark is-rounded">
69+
<span class="icon"><i class="fas fa-star"></i></span>
70+
<span>OpenReview</span>
71+
</a>
72+
<a href="https://github.com/KGML-lab/Generalized-Forward-Inverse-Framework-for-DL4SI" class="button is-dark is-rounded">
73+
<span class="icon"><i class="fab fa-github"></i></span>
74+
<span>Code</span>
75+
</a>
76+
<a href="https://iclr.cc/virtual/2025/poster/27736" class="button is-dark is-rounded">
77+
<span class="icon"><i class="fas fa-file-pdf"></i></span>
78+
<span>PDF</span>
79+
</a>
80+
</div>
81+
</div>
82+
</div>
83+
</section>
84+
85+
<section class="section">
86+
<div class="gfi_framework">
87+
<figure>
88+
<img src="./static/images/GFI_Framework.png" alt="GFI Framework" loading="lazy" width=25%>
89+
<figcaption> Figure 1: A unified framework for solving forward and inverse problems in subsurface imaging. </figcaption>
90+
</figure>
91+
</div>
92+
<hr style="width: 60%; margin: 2rem auto;">
93+
</section>
94+
95+
<!-- Abstract -->
96+
<section class="section">
97+
<div class="container is-max-desktop has-text-justified">
98+
<h2 class="title is-3"> Abstract </h2>
99+
<p>
100+
In subsurface imaging, learning the mapping from velocity maps to seismic waveforms (forward problem) and waveforms to velocity (inverse problem) is important for several applications. While traditional techniques for solving forward and inverse problems are computationally prohibitive, there is a growing interest in leveraging recent advances in deep learning to learn the mapping between velocity maps and seismic waveform images directly from data.
101+
Despite the variety of architectures explored in previous works, several open questions remain unanswered such as the effect of latent space sizes, the importance of manifold learning, the complexity of translation models, and the value of jointly solving forward and inverse problems.
102+
We propose a unified framework to systematically characterize prior research in this area termed the Generalized Forward-Inverse (GFI) framework, building on the assumption of manifolds and latent space translations.
103+
We show that GFI encompasses previous works in deep learning for subsurface imaging, which can be viewed as specific instantiations of GFI.
104+
We also propose two new model architectures within the framework of GFI: Latent U-Net and Invertible X-Net, leveraging the power of U-Nets for domain translation and the ability of IU-Nets to simultaneously learn forward and inverse translations, respectively.
105+
We show that our proposed models achieve state-of-the-art performance for forward and inverse problems on a wide range of synthetic datasets and also investigate their zero-shot effectiveness on two real-world-like datasets.
106+
</p>
107+
</div>
108+
<hr style="width: 60%; margin: 2rem auto;">
109+
</section>
110+
111+
<section class="section">
112+
<div class="container is-max-desktop has-text-justified">
113+
<h2 class="title is-3">Method Overview</h2>
114+
<p>
115+
We propose Generalized Forward-Inverse (GFI) framework based on two assumptions. First, according to the manifold assumption, we assume that the velocity maps v ∈ v&#119985; and seismic
116+
waveforms p ∈ p&#119985; can be projected to their corresponding latent space representations, v&#771; and p&#771;, respectively, which can be mapped back to their reconstructions in the original space, v&#770; and p&#770;.
117+
Note that the sizes of the latent spaces can be smaller or larger than the original spaces. Further, the size of v&#771; may not match with the size of p&#771;. Second, according to the latent space
118+
translation assumption, we assume that the problem of learning forward and inverse mappings in the original spaces of velocity and waveforms can be reformulated as learning translations in their
119+
latent spaces.
120+
</p>
121+
<ol>
122+
<li>
123+
<b>Latent U-Net Architecture: </b> We propose a novel architecture to solve forward and inverse problems using two latent space translation models implemented using U-Nets, termed Latent U-Net. Latent
124+
U-Net uses ConvNet backbones for both encoder-decoder pairs: <code>E</code><sub>v</sub>, <code>D</code><sub>v</sub> and <code>E</code><sub>p</sub>, <code>D</code><sub>p</sub>, to project
125+
v and p to lower-dimensional representations. We also constrain the sizes of the latent spaces of
126+
v&#771; and p&#771;to be identical, i.e., dim(v&#771;) = dim(p&#771;), so that we can train two separate U-Net models to implement the latent space mappings L<sub>v&#771; &rarr; p&#771;</sub> and L<sub>p&#771; &rarr; v&#771;</sub>.
127+
128+
<div class="latent_unet">
129+
<figure>
130+
<img src="./static/images/LatentU-Net.png" alt="Latent U-Net architecture" loading="lazy" width=45%>
131+
<figcaption> Latent U-Net architecture </figcaption>
132+
</figure>
133+
</div>
134+
135+
</li>
136+
<li>
137+
<b>Invertible X-Net Architecture: </b> We propose another novel architecture within the GFI framework termed Invertible X-Net to answer
138+
the question: “can we learn a single latent space translation model that can simultaneously solve
139+
both forward and inverse problems?” We employ invertible U-Net in the latent spaces of velocity and waveforms, which can be constrained to be of the same size (just like Latent-UNets), i.e.,
140+
dim(v&#771;) = dim(p&#771;).
141+
<div class="inv_xnet">
142+
<figure>
143+
<img src="./static/images/InvertibleX-Net.png" alt="Invertible X-Net architecture" loading="lazy" width=45%>
144+
<figcaption> Invertible X-Net architecture </figcaption>
145+
</figure>
146+
</div>
147+
</li>
148+
</ol>
149+
</div>
150+
<hr style="width: 60%; margin: 2rem auto;">
151+
</section>
152+
153+
<!-- Results Table Placeholder -->
154+
<section class="section">
155+
<div class="container is-max-desktop">
156+
<h2 class="title is-3">Experiments</h2>
157+
</div>
158+
</section>
159+
160+
<!-- BibTeX -->
161+
<section class="section" id="bibtex">
162+
<div class="container is-max-desktop content">
163+
<h2 class="title">BibTeX</h2>
164+
<pre><code>@inproceedings{
165+
gupta2025a,
166+
title={A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations},
167+
author={Naveen Gupta and Medha Sawhney and Arka Daw and Youzuo Lin and Anuj Karpatne},
168+
booktitle={The Thirteenth International Conference on Learning Representations},
169+
year={2025},
170+
url={https://openreview.net/forum?id=yIlyHJdYV3}
171+
}</code></pre>
172+
</div>
173+
</section>
174+
175+
<section class="section" id="acknowledgement">
176+
<div class="container is-max-desktop content">
177+
<h2 class="title">Acknowledgement</h2>
178+
<p>
179+
This work was supported in part by NSF awards IIS-2239328 and IIS-2107332. We are grateful
180+
to the Advanced Research Computing (ARC) Center at Virginia Tech for providing access to GPU
181+
compute resources for this project. This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government
182+
retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the
183+
published form of this manuscript, or allow others to do so, for US government purposes. DOE will
184+
provide public access to these results of federally sponsored research in accordance with the DOE
185+
Public Access Plan ( https://www.energy.gov/doe-public-access-plan).
186+
</p>
187+
</div>
188+
</section>
189+
</main>
188190

189191
</body>
190192
</html>

0 commit comments

Comments
 (0)