Some thoughts regarding general AI conferences
Chaopeng Shen
Dear all,
After I returned from the ICLR AI conference last month and finished hydroML early in late May, I have gathered some insights that I’d like to share. I think generic AI conferences are good opportunities to generate cross-domain impacts for your work, and also a way to demonstrate that our domain insights offer value to the greater AI communities. Let's just say --- I have seen very few geoscientific/hydrologic colleagues here. There was only some climate-focused work but they came from the AI community or industry.
The AI community as a whole places great importance on peer-reviewed conferences, rather than journal publications. Many foundational developments are documented only via conference papers, making these conferences critical for achieving recognition and impact in the field.
NIPS, ICLR, & ICML are regarded as the best 3 generic AI conferences that have published high-impact and extremely-highly-cited papers like "Attention is all you need" (NIPS2017, cited 181,436 times as I just checked), "Denoising Diffusion Probabilistic Models" (NIPS2015, cited 21,118 times), “Auto-Encoding Variational Bayes” (ICLR2013, cited 44,578 times), and "Fourier Neural Operators" (ICLR2020, cited 3,050 times). I list these not to say we should pursue high citations for the sake of citations, but to suggest these conferences as good opportunities to expand the impacts of your work to a larger audience. AAAI and KDD are also great, among others. Here is a ranking if you are interested. Their proceedings website typically does not look that formal, though.
Overall, I want to encourage members of the hydrology community to submit more foundational methodological advances here, especially those inspired by our domain insights, because that makes the work unique.
If you are interested in learning more:
1. These conferences tend to have two tracks: one main "conference" and many "workshops". Papers accepted to main conferences are collected into formal proceedings which are SCI-indexed (but they use crappy websites and rely on openreview or arxiv for archiving). In contrast, workshops are set up more like AGU-style sessions but with short-format online papers (most likely not indexed or as rigorously peer reviewed as the main conference). With respect to on-site presentations, most papers are assigned to posters, with oral presentations likely those with high peer-review scores.
2. AI conferences now attract people from all over the scientific domains, meaning you can generate impact outside of your typical domain. Reviewers could be from computer science,, applied maths, biomedical, mechanical, or electrical engineering. Despite the disparate domains, people speak the same ML language (almost like soccer or music). Reviewers tend to prefer core methodological advances relevant to many domains, with a ton of irrefutable benchmarks. Many math problems are similar at the core.
3. The review process is rigorous, cross-disciplinary and sometimes ferocious, featuring double-blind (unless you arXiv it first) open reviews with comments and responses publicly available for all to see. Each submission is assigned 4-5 reviewers who are often not familiar with domain materials, so you need to write in generic terms. My general experience, however, is that if you have truly innovative and foundational work and explain it well in a domain-agnostic manner, reviewers can understand and acknowledge it.
4. The conferences have fast review times (1 month initial review; 2.5 months total including rebuttal periods). But there is a need for reciprocal review service. You have to perform review services for 4-6 papers.
5. There is also work on theoretical analysis, deep insights about algorithm behaviors, and those who are in trendy topics (like LLMs these days) may have a higher chance of getting accepted, but innovative scientific ML is also accepted in top-3 conferences. Applications of existing ML algorithms are not often viewed favorably, it seems. It seems many papers propose foundational-level progress although I cannot say this would always be the case for the main AI papers (simply because I’m not in the LLM business).
6. As an example, in our ICLR 2025 paper (https://arxiv.org/abs/2505.08740) we show that AI and differentiable approaches are applicable to solving high-resolution environmental partial differential equations (PDEs). This work actually came from a domain insight when working on our past differentiable parameter learning and differentiable modeling work: parameters are super important in our models but surrogate models often fail to build the correct internal relationships, and these relationships/sensitivities are crucial for extrapolation. In general, ML models can sometimes give you the right results for the wrong reasons. Recently proposed AI-infused PDE solvers (or neural operators) are incredibly efficient, but often suffer from incorrect learned relationships and sensitivities, resulting in large uncertainty in inversion and optimization tasks. We thus proposed Sensitivity-Constrained Fourier Neural Operators (SC-FNO), which train on not only the solution but also the sensitivities (Jacobians). SC-FNO shows robustness even under sparse training data or concept drift scenarios, permitting accurate inversions. We mostly used general PDEs like Burger's and Navier Stokes as examples and provided lots of benchmarks. This paper received ferocious reviews but, in the end, reviewers understood the work and accepted it.
The above are my thoughts. Feedback welcomed.
Thanks and best.
The Pennsylvania State University
University Park, PA 16802
Office: 814-863-5844
Solomon Vimal
Dear Prof. Shen and the Hydrology Community,
Prof. Shen’s takeaway messages from ICRL and HydroML were deeply thought-provoking (I attended HydroML as well), and I feel compelled to share a few reflections. My message does not contradict Prof. Shen’s, but complements it. I fully support his leadership in advancing AI within hydrology and see his work as a model of “physics-informed AI” done right.
I identify strongly with both the hydrology and AI communities, though my path has since diverged into industry and AI entrepreneurship. I co-founded my first AI startup in 2016 alongside the now-Director of AI at Google, who currently leads Google’s Gemini Mini Foundational Model. Today, I’m building a new AI company (geothara.com) rooted in a decade of experience at the intersection of advanced AI, hydrology, and climate technology. After reading Prof. Shen's email, I wrote a 3-page essay outlining my perspective on the use of AI in hydrology, especially contrasting academia and industry approaches. If you’re interested in reading it and engaging with this work, you can access the full essay here or feel free to reach out to me directly. Here’s a brief summary:
Key Takeaways:
- AI Literacy First: Before most of our community can work at the level Prof. Shen exemplifies, we need to master AI literacy, including the art of prompting, not just modeling.
- Return to Fundamentals: We must reconnect with the foundational physics of water. I encourage everyone to watch this fun short clip of Richard Feynman to embody this spirit of getting the physics right at all scales.
- Read Horton, Seriously: Our field had its Feynman in Robert E. Horton. Over 200+ and 300+ books worth of materials by him have never been seen or cited; you can find a full list, and the backstory (in a draft blog) on my site: https://geothara.com.
- AI Can Help Us: At Geothara we are developing tools to use modern AI (like a Hydro-LLaMA) to revisit and make sense of Horton’s legacy. If this interests you, please contact me to collaborate.
- Caution: If we don’t read foundational works deeply, like Prof. Shen’s early DL papers (WRR 2014 onwards) and Horton’s original work, our community risks becoming a case study in “physics-uninformed AI.”
Thank you for your time, and for being part of a field I care deeply about. And thank you again to Prof. Shen for inspiring this important conversation and continuing to lead by example.
Best regards,
Solomon Vimal
Founder and CEO, Geothara
https://geothara.com
--
Send an email to the whole group by emailing hydrom...@hydroml.org. Contact us or ask a moderator to post an anonymous question for you by contacting in...@hydroml.org. To unsubscribe from this group, send an email to hydroml.grou...@hydroml.org.
To unsubscribe from this group and stop receiving emails from it, send an email to hydroml.grou...@hydroml.org.
Amobichukwu Amanambu
Dear Dr Shen,
Thank you for taking the time to share your experiences and insights from ICLR and hydroML. Although I haven’t yet attended these conferences myself, your message gave me a clear and honest window into how foundational and impactful AI research can be—especially when it’s driven by real domain understanding like in hydrology.
One thing that stood out to me was how you described the value of domain insights in shaping core methodological advances. It almost feels like AI, when used thoughtfully, brings us closer to human-like reasoning—especially when we work to ensure models not only produce the right answers, but for the right reasons. That perspective really resonates with me.
I’m excited to explore how our field can contribute more actively to this broader scientific dialogue. Your example of SC-FNO also helps show what’s possible when we bridge physical understanding with methodological innovation.
Thanks again for the thoughtful encouragement. It’s inspiring to see how people in our community are helping shape the future of AI from a hydrologic perspective.
Warm regards,
Amobi
--
For see all the Job and Positions give a look at: https://hydrojobs.blogspot.com/
To see the webinars announced: https://hydrowebinartribune.blogspot.com/
---
You received this message because you are subscribed to the Google Groups "AboutHydrology" group.
To unsubscribe from this group and stop receiving emails from it, send an email to abouthydrolog...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/abouthydrology/CAFqLm1qpyM-KtTG%3DSbF6f1HrQUQqOt%3DGb2Q2qr_phEASRzzk1A%40mail.gmail.com.