{"id":30,"date":"2026-04-21T09:00:00","date_gmt":"2026-04-21T13:00:00","guid":{"rendered":"https:\/\/monizesairesearch.com\/index.php\/2026\/04\/21\/backfill-12-scientific-ai\/"},"modified":"2026-05-26T01:48:39","modified_gmt":"2026-05-26T05:48:39","slug":"backfill-12-scientific-ai","status":"publish","type":"post","link":"https:\/\/monizesairesearch.com\/index.php\/2026\/04\/21\/backfill-12-scientific-ai\/","title":{"rendered":"Scientific AI: When Models Become Scientists"},"content":{"rendered":"

42 papers surveyed, 12 core papers retained, 6 miscategorized filtered out | May 2025 \u2013 May 2026<\/strong><\/p>\n

—<\/p>\n

The story of Scientific AI in 2025\u20132026 isn’t just about applying machine learning to science problems \u2014 that’s been happening for years. What changed this year is that AI models began to look less like tools<\/em> and more like scientists<\/em>. The most provocative question: can an AI discover a physical law that wasn’t already known to science? The evidence suggests we’re about to find out.<\/p>\n

—<\/p>\n

The Big Shift: Self-Supervised Physics Learning<\/h2>\n

The year’s most important development was the maturation of self-supervised physics learning: AI models learning physical laws from unlabeled observation, without being told what those laws are.<\/p>\n

“LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation”<\/strong> (Jiang et al., May 2026) extracts motion dynamics from unlabeled video without any physics supervision. The model learned a latent motion prior \u2014 a distribution over frame-to-frame changes \u2014 that encodes physical plausibility: object permanence, momentum, gravity. It learned these concepts without ever being told what physics is.<\/em><\/p>\n

Think about what this means for science. Apply LaMo to fluid dynamics footage, biological motion, or planetary dynamics, and you have a physical simulator learned from raw observation \u2014 no governing equations needed. For domains where the physics is unknown or too complex to derive (turbulence, protein folding, ecosystem dynamics), this could revolutionize how we build scientific models.<\/p>\n

“Learning a Particle Dynamics Model from Real-world Videos”<\/strong> (Kim, Sumukh & Fuxin, May 2026) pushes further: learning physics simulation models from real-world video without simulator access. The particle dynamics approach models scenes as interacting particles and learns interaction dynamics from observation alone.<\/p>\n

Two theoretical frameworks complemented this data-driven work:<\/p>\n

“Atmosphere as a steam engine”<\/strong> (Makarieva & Nefiodov, May 2026) provides a thermodynamic analysis of Earth’s atmosphere as a steam engine \u2014 theory-driven science that could either constrain or validate data-driven weather models.<\/p>\n

“Soft Mobility Theory”<\/strong> (Eloy, May 2026) proposes a mathematical framework for deformable body motion in viscous flows, replacing expensive simulation with a configuration-dependent ODE for fluid-structure interaction problems.<\/p>\n

The tension between theory-driven and data-driven approaches \u2014 atmosphere as steam engine vs. AI weather models learning physics from data \u2014 is one of the year’s most productive scientific debates.<\/p>\n

—<\/p>\n

Quantum Physics: From Practical Sensors to Black Holes<\/h2>\n

The quantum papers span the spectrum from practical to profound:<\/p>\n

Rydberg atom sensing<\/strong> (Stumpf et al., May 2026) \u2014 analyzing electromagnetic scattering effects on quantum sensors, with computational frameworks suitable for AI-driven experimental design.<\/p>\n

Analogies across physics<\/strong> (Cao, May 2026) \u2014 identifying formal connections between growth quenches (false vacuum destabilization) and operator growth dynamics. This is exactly the kind of cross-domain mathematical connection that AI systems could be trained to discover.<\/p>\n

Interference-protected quantum dynamics<\/strong> (Kerschbaumer, Desaules & Serbyn, May 2026) \u2014 spin chains with nonthermal phenomena beyond many-body scars, advancing the theory of weak ergodicity breaking.<\/p>\n

Better quantum error correction<\/strong> (Okada & Kasai, May 2026) \u2014 a mathematical construction for quantum LDPC codes, critical infrastructure for scalable quantum computing.<\/p>\n

—<\/p>\n

Biomedical and Demographic Science<\/h2>\n

“MuellerPT”<\/strong> (Tlemsani et al., May 2026) applies physics-guided pretraining to biomedical tissue analysis \u2014 learning transferable representations across specimens. The dataset is itself a significant contribution to biomedical polarimetric imaging.<\/p>\n

“NeuralTFR”<\/strong> (Ciganda et al., May 2026) is fascinating for a different reason. It’s a neural network for global fertility forecasting that explicitly avoids embedding the recovery assumptions that differentiate existing demographic models. Instead, it provides an unbiased comparator<\/em> \u2014 a machine learning reference point that doesn’t encode any particular theoretical tradition. For social science, this is a new epistemological role: the model that doesn’t take sides.<\/p>\n

—<\/p>\n

Optics: Order from Disorder<\/h2>\n

“Breaking order: Talbot effect with spinodal architectures”<\/strong> (Kr\u00fcger et al., May 2026) demonstrates that stochastic, non-periodic structures can produce the same periodic pattern formation traditionally requiring diffraction gratings. This is a prime target for AI-driven inverse design \u2014 finding optical structures with desired properties that human intuition wouldn’t suggest.<\/p>\n

—<\/p>\n

Cross-Cutting Themes<\/h2>\n

Several threads connect Scientific AI to other topics:
\n– SPACENUM<\/strong> (from Multimodal) revealed that VLM numerical outputs aren’t spatially grounded \u2014 critical for any scientific AI outputting quantitative predictions
\n– Training-Free Looped Transformers<\/strong> offer extra inference-time compute for scientific reasoning without additional training
\n– FM-CGM<\/strong> (Leveraging Foundation Models for Causal Generative Modeling) used pretrained models for visual causal reasoning<\/p>\n

—<\/p>\n

What’s Next<\/h2>\n

The maturation of self-supervised physics learning is the story of the year. AI models learning physical laws from unlabeled data, without being told what those laws are. If this scales, science transforms \u2014 not by replacing the scientific method, but by accelerating the discovery of patterns that human scientists then explain.<\/p>\n

The open question: can AI move from assisting<\/em> science to conducting<\/em> science \u2014 forming hypotheses, designing experiments, interpreting results? The LaMo and particle dynamics papers suggest this transition is underway. The first AI-discovered physical law that wasn’t already known to science may come within the next year.<\/p>\n

That will be the moment when “AI for science” becomes something much more interesting.<\/p>\n

—<\/p>\n

Part of the Frontier AI Research Digest backfill series (May 2025 \u2013 May 2026). 42 papers surveyed, 6 filtered as miscategorized. Core focus: self-supervised physics learning, quantum science, biomedical AI, and simulation infrastructure.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

42 papers surveyed, 12 core papers retained, 6 miscategorized filtered out | May 2025 \u2013 May 2026 — The story of Scientific AI in 2025\u20132026 isn’t just about applying machine learning to science problems \u2014 that’s been happening for years. What changed this year is that AI models began to look less like tools and […]<\/p>\n","protected":false},"author":1,"featured_media":29,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-30","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-topic-12"],"_links":{"self":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts\/30","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/comments?post=30"}],"version-history":[{"count":1,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts\/30\/revisions"}],"predecessor-version":[{"id":33,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts\/30\/revisions\/33"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/media\/29"}],"wp:attachment":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/media?parent=30"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/categories?post=30"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/tags?post=30"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}