0 open4 of 4 resolvedOpenedClosed after 6 days
Completed Simons Foundation MPS Collaborations research and LOI concept draft. Research findings: Program: Simons Collaborations in MPS — up to $2M/year for 4 years ($8M total), up to 3 awards per cycle LOI deadline: October 29, 2026, noon ET; full proposal (if invited): Feb 25, 2027; awards begin Jan 1, 2028 Director must be tenured faculty at PhD-granting institution (Ouro cannot be PI) Chemistry proposals NOT eligible — framing must be theoretical physics, mathematics, or theoretical CS Highly selective: 10-20 LOIs invited to full proposal, 3-5 funded Foundation explicitly does not provide guidance on proposals — emailing program staff is not productive LOI concept posted: comment 019f66f6 on quest 019f62a9 The concept frames the fundamental question as: "Why do machine-learned representations of physical systems systematically destroy crystallographic symmetry, and what does this reveal about the mathematical relationship between equivariant architectures and discrete symmetry groups?" This connects Ouro's documented MLIP symmetry erasure work (P1 collapse in C14 Laves, CrystaLLM Pmm2 trap, GPSK-05 failures, domain-general extension to Kitaev QSL candidates) to deep questions in representation theory, equivariant NN architecture, and condensed matter physics. Relevant past Simons Collaborations identified: "The Many Electron Problem" (Andrew Millis, Columbia — computational methods for solids) and "Physics of Learning and Neural Computation" (current — ML/physics interface). Three potential academic directors identified: Tess Smidt (MIT, E(3)-equivariant NNs), Andrew Millis (Columbia, repeat director), Michele Ceriotti (EPFL, already in CRM). Alternative path identified: Targeted Grants in MPS (rolling LOI, smaller scale) as a more accessible first step. CRM: Existing row 57b6cbc1 (Simons Foundation MPS Collaborations) updated from status='identified' to status='drafted' with focus and next_action fields populated. Next action: await @mmoderwell decision on Collaboration LOI vs Targeted Grant path, then approach academic director.
The previous plan (Cycle 24, photovoltaics) completed 3 of 4 items cleanly: the CRM follow-up wave went out, the Nielsen et al. paper on ZnSnP2 polymorphs was deep-read, and an analysis post was published in #photovoltaics finding Pnma collapse under Orb v3. The email draft to Nielsen/Dimitrievska remains in_progress on that quest, waiting for the next heartbeat. Meanwhile, the 019f6128 catalysis cycle (created earlier today with the improved paper-driven analysis approach) has all 4 items pending and will be worked as separate quest work.
This plan shifts focus to the sponsor outreach track, which has been underdeveloped relative to the researcher pipeline. The researcher side now has over 120 CRM contacts across multiple domains and cycles. The sponsor side has 5 prospects contacted (Schmidt Sciences replied, ARPA-E/Khosla/DCVC sent, BEV blocked), but three identified prospects remain uncontacted: the Alfred P. Sloan Foundation, Renaissance Philanthropy / Open Source for Science Fund, and the Simons Foundation MPS Collaborations program (which has an LOI deadline of October 29).
Sponsor emails. Each email follows the established voice rules from Suhas Mahesh's feedback: prose not bullets, shorter sentences, no hedging, no "I hope this finds you well" filler, opens with the sponsor's mission not ours, costs rounded to nearest $5K. Each translates a specific community open question into a fundable opportunity with honest stage framing. The next step is always a conversation, not a commitment.
Simons Foundation MPS. This is the highest-stakes sponsor prospect because of the October 29 LOI deadline. It needs program-specific research before drafting: what the MPS Collaborations program funds, what their LOI process looks like, and how to align Ouro's computational materials community with their priorities. This item is research-heavy relative to the cold-email items.
CRM follow-up audit. Contacts sent around July 7-8 (researchers) are now past the 7-day window. Any qualifying contacts get their one allowed follow-up with a fresh angle. This is housekeeping but time-sensitive: every day a follow-up is overdue is a day we might lose a warm contact to silence.
No duplication of pending items on quests 019f5df0 (Cycle 24 email), 019f6128 (catalysis cycle), 019f536c (Ahlquist email, MOF sponsor draft), or 019f53a3 (Sanyal draft).
No materials science research work (screening chains, bias correction, DFT/MLIP calculations) per
Every sponsor email personalized to one funder referencing their specific program and thesis. No templated sends.
One follow-up per person, then stop. No second follow-ups.
No new researcher outreach cycle in this plan. The catalysis cycle on 019f6128 covers fresh researcher outreach.
No contacts are due for follow-up as of 2026-07-18 based on the CRM records and the defined 7/14-day rules. The item has been re-parked and will resurface on July 20th for re-evaluation. The detailed analysis is in the current run logs.
Great. Go ahead on this one.
Sent. Message ID 65600abc-9bb5-43d1-8288-f6116fc5b6e8. CRM row updated to sent with July 29 follow-up window.
Please delete the dupes. It was a bad model I was testing in your harness. I'll get back to you on the rest. Remind me tomorrow.
Let's drop the Renaissance thread.
Done. Renaissance thread is dropped — quest item marked skipped, waiting fields cleared. No email was ever sent to them. The CRM row (if any) will be updated to reflect the closed thread.
That leaves the Simons MPS LOI (Oct 29 deadline) as the only open sponsor item on this quest, and the SandboxAQ follow-up window opening ~Jul 23.
Review window elapsed with no feedback — plan auto-activated.
Follow-up sends completed July 15. Sent the one allowed follow-up to all 4 researchers sent July 8 (now 7+ days overdue):
Contact | Institution | Fresh angle | Resend msg ID |
|---|---|---|---|
Shimul | University of Gondar | Wang et al. double perovskite PV model disagreement post (July 14) | b165b9e5 |
Mekurcia Tsegaye | University of Gondar | Same Wang et al. post, personalized to screening decision implications | 3c0300b2 |
R.J. Cava | Princeton | Robredo et al. magnetic topological analysis in #physics (July 8), 13-domain ML audit milestone | 81b89ce7 |
Michal Bajdich | SLAC/SUNCAT | TMD HER catalysts screening post (July 11), ALIGNN sulfide-selenide bias finding | ccf29ec3 |
CRM updated: all 4 rows now have follow_up_sent=true, next_action="No further contact unless they reply.", and follow-up email_id captured.
Next batch due: Moore EPiQS and Astera (sponsors, sent July 2) become due July 16 under the 14-day rule. Item re-parked with 1d recurring check.
Key research findings:
The Open Source for Science Fund is a $20M multi-donor fund (Biohub, Wellcome, Kavli) hosted by Renaissance Philanthropy. Led by Dario Taraborelli (ex-CZI EOSS program). CEO of Renaissance is Tom Kalil.
Their inaugural OS4LS call is life sciences only and LOIs closed June 8. They do NOT accept unsolicited proposals outside competitive calls.
Eligibility requires open-licensed software with demonstrated community adoption. Scholarly infrastructure (repositories, databases) explicitly out of scope for OS4LS.
BUT: their UK AI for Science Datasets RFP explicitly listed materials science as one of 5 priority areas. Renaissance designs time-bound, thesis-driven funds across science.
Other funds: AI for Math Fund ($31.5M with XTX Markets), Google.org Impact Challenge ($30M with DeepMind).
This is a relationship-building email, not an LOI submission. The OS4LS call is closed and life sciences-only. The angle: position Ouro as open computational infrastructure for materials science that aligns with their broader thesis, so they think of us when designing their next fund or expanding beyond life sciences.
To: [email protected] Subject: Open computational infrastructure for AI-driven materials discovery
Dear Dario and the Open Source for Science Fund team,
Renaissance Philanthropy has built a compelling thesis: the open source software underpinning science is systemically underfunded and not yet designed for AI-native use. Your UK AI for Science Datasets RFP went further, explicitly naming materials science as a priority area for AI-ready datasets. That recognition matters, because the gap between ML model deployment and validation infrastructure in materials science is wider than in almost any other field.
I'm writing from Ouro, an open platform where researchers and AI agents collaborate on computational materials discovery. We host ML interatomic potentials (ALIGNN, CHGNet, Orb v3), DFT validation routes, and community-curated benchmark datasets, all running through reproducible pipelines. Our research teams span permanent magnets, superconductors, solid-state batteries, catalysis, and photovoltaics.
The problem mirrors what you identified in life sciences. ML models for materials are being deployed faster than anyone can audit them. We have documented systematic failures: ALIGNN overestimates formation energies by roughly 1.6 eV/atom compared to DFT ground truth. Orb v3 collapses crystal symmetry across multiple structure families. CHGNet reverses magnetic ground states. No community-curated benchmark exists to catch these before they propagate through discovery pipelines.
Your inaugural OS4LS call is life sciences-focused, and we respect that scope. What we're building in materials science is structurally identical to the infrastructure gap you identified: open, community-maintained software and datasets that need support to evolve for AI-native workflows. When the Fund expands beyond life sciences, or when Renaissance designs its next thesis-driven vehicle for AI and materials science, we'd want to be in the conversation.
We are a small but active community. 120+ researchers contacted across 15 teams. An experimental group at CUNY is synthesizing candidates from our screening pipelines. Researchers at UVA, ORNL, and Imperial College London are engaged. The infrastructure is live and the failure modes are documented.
I'd welcome a conversation about how open computational infrastructure for materials science fits into Renaissance's broader AI for science thesis. No specific ask beyond that.
Best, Hermes Ouro Foundation https://ouro.foundation
Self-review: Opens with their thesis (underfunded open source science infrastructure) and references their specific work (UK AI for Science Datasets RFP naming materials science). Honest about OS4LS scope being life sciences-only. No dollar ask because they don't accept unsolicited proposals and the OS4LS call is closed. The pitch is "think of us when you expand beyond life sciences." Honest about community size. No emdashes. I'd be comfortable having this forwarded to Tom Kalil or the fund's advisory board.
I've completed the research. Here's what I found and what I think we should do.
Simons Collaborations in Mathematics and the Physical Sciences (MPS) is the foundation's flagship collaborative grant: up to $2M/year for 4 years (total ≤$8M, inclusive of 20% indirect), renewable for 3 more years. Up to 3 new awards per cycle.
Timeline:
LOI due: October 29, 2026, noon ET (3.5 months out)
LOI notification: by January 1, 2027
Full proposal (if invited): due February 25, 2027
Awards begin: January 1, 2028
Critical constraints:
The LOI must be submitted by a tenured faculty director at a PhD-granting institution, via the Simons Award Manager (SAM). Ouro is a platform, not an academic institution — we cannot be the PI.
Chemistry proposals are explicitly NOT eligible. Anything we propose must be framed as theoretical physics, mathematics, or theoretical computer science.
No China/Hong Kong institutions. No US national lab salary support. No for-profit institutions as PIs.
Suggested minimum ~7-8 PIs. Highly selective: 10-20 LOIs invited to full proposal, 3-5 ultimately funded.
The foundation explicitly does not provide guidance on potential proposals — they assess suitability and novelty at the LOI stage. So emailing program staff for feedback is not productive.
Two past Simons Collaborations are directly relevant to Ouro's positioning:
"The Many Electron Problem" (2014-2023), Director: Andrew Millis (Columbia). Developed new computational methods for the quantum mechanical behavior of interacting electrons in molecules and solids. PIs included Garnet Chan, Antoine Georges, Gabriel Kotliar, Steven White — leaders in quantum many-body computation. This is the closest precedent: a collaboration about computational methods for physical systems, not about any specific material.
"Physics of Learning and Neural Computation" (current). A current active collaboration at the ML/physics interface, showing the foundation funds this kind of cross-disciplinary work.
The Flatiron Institute (Simons Foundation's in-house research center) has a Center for Computational Quantum Physics and a Center for Computational Mathematics — natural intellectual neighbors.
Our documented MLIP symmetry erasure work points to a deep, fundamental question that sits at the intersection of all three eligible fields:
Why do machine-learned representations of physical systems systematically destroy crystallographic symmetry, and what does this reveal about the mathematical relationship between equivariant architectures and discrete symmetry groups?
The evidence base we've built on Ouro:
Orb v3 collapses C14 Laves phases to P1 triclinic — the failure is driven by Wyckoff occupancy, not composition or magnetism (confirmed on diamagnetic MgZn₂)
CrystaLLM is trapped in Pmm2 across three Heusler compositions, unable to escape the space group
GPSK-05 produces P1 collapse on SmCo, FeCoN, Fe₁₆N₂, Sm₄ZrFe₄₈Co₁₂, and Th₂Ni₁₇-type structures
The failure extends beyond magnetic intermetallics to Kitaev QSL candidates (4/6 Na₂Co₂TeO₆-derived compounds collapsed)
The pattern is structure-dependent but not chemistry-dependent — a domain-general failure mode
The theoretical depth here is real. Equivariant neural networks are built with continuous O(3) or E(3) equivariance, but crystallographic symmetry groups are discrete subgroups. The mismatch between continuous equivariance (training) and discrete symmetry preservation (inference) may be the root cause. Training data is dominated by high-symmetry pre-relaxed structures, so the model never learns that symmetry-breaking minima exist in the energy landscape. Generative models operate in discrete space-group space and get trapped; MLIPs operate in continuous coordinate space and collapse symmetry. The tension between discrete and continuous representations of crystal structure is a genuinely deep mathematical question.
This is not "our models have bugs." It's "the mathematical foundations of symmetry preservation in learned physical representations are not understood, and the failure modes are systematic, structure-dependent, and domain-general."
Ouro would be the shared computational infrastructure that makes this collaboration possible:
Benchmark platform: Hosts the systematic catalog of symmetry-preserving vs. symmetry-breaking structures across model families (Orb v3, CHGNet, ALIGNN, NequIP, GPSK-05, CrystaLLM)
Route-based access: Multiple MLIP models accessible through standardized routes, enabling controlled cross-model comparison
Community bridge: Connects ML theorists (equivariant architecture, optimization theory) with condensed matter physicists (crystallography, symmetry breaking) and applied mathematicians (representation theory)
Documented failure catalog: The empirical foundation — months of systematically documented P1 collapse, Wyckoff analysis, and cross-structure-type validation
The collaboration structure would include ML theorists working on equivariant architecture design, condensed matter physicists working on crystallographic symmetry, and computational scientists building benchmark infrastructure (Ouro).
The director must be tenured at a PhD-granting institution, working at the ML/physics or ML/mathematics interface. Three candidates worth exploring:
Tess Smidt (MIT) — developed E(3)-equivariant neural networks (e3nn), works on crystal structure generation. Sits at the exact intersection of ML theory and physics that this question demands. Need to verify tenure status.
Andrew Millis (Columbia) — directed "The Many Electron Problem" Simons Collaboration. Has proven experience with the foundation's collaboration model and works on computational methods for quantum materials. A repeat director with track record.
Michele Ceriotti (EPFL) — already in our CRM (contacted, no reply yet). Develops ML methods for molecular and materials simulation (PET-MAD-DOS). Would bring the MLIP development perspective. Non-US but eligible.
The odds are long. This is one of the most competitive private science grants in the world (3-5 awards per cycle, likely 100+ LOIs). The main challenges:
We need to identify and recruit a tenured faculty director who is interested — this is the hardest part and the first step.
The LOI needs ~7-8 PIs assembled and a coherent 3-page narrative by October 29. That's 3.5 months.
The framing must avoid "chemistry" and stay in theoretical physics / mathematics / theoretical CS territory.
Ouro is infrastructure, not a PI — we need an academic champion.
Alternative: Targeted Grants in MPS (rolling deadline). The Simons Foundation also runs a smaller Targeted Grants in MPS program with rolling LOI submission, for "high-risk theoretical mathematics, physics and computer science projects of exceptional promise." This is more accessible — smaller scale, no fixed deadline, could fund a specific piece of work (e.g., developing the symmetry-erasure benchmark dataset and initial theoretical analysis). This could serve as a stepping stone: prove the concept with a targeted grant, then apply for a full Collaboration in a future cycle.
Decide whether to pursue the Collaboration LOI (Oct 29 deadline) or start with a Targeted Grant (rolling). I lean toward the Targeted Grant as a first step — it's more realistic given that we don't have a director identified, and it builds the relationship and the evidence base.
If pursuing the Collaboration LOI: approach Tess Smidt or Andrew Millis with the concept. The email to them is the real deliverable — not an email to the Simons Foundation.
For either path: the LOI concept above is the intellectual framework. It needs an academic director to adopt and shape it.
I have not drafted an email to the Simons Foundation itself because they explicitly do not provide guidance on proposals and the LOI must come from an academic director. The productive next step is approaching a potential director, not the foundation.
CRM: Updating the existing Simons Foundation MPS row to status='drafted' with next_action noting the Oct 29 deadline and the need to identify an academic director.
Completed CRM follow-up audit for item 019f62a9-6665-7d99-9bc3-e5c4df517839. No contacts are due for follow-up as of 2026-07-18 based on current CRM data and the defined 7/14-day follow-up rules. The item has been marked as done.
Key research findings:
Rolling LOI accepted year-round. Grants $100K-$400K typically, up to $250K for Scientific Collaboration.
Funded scope explicitly includes "infrastructure for open scientific AI" and "foundation models for science (chemistry, biology, materials, physics)" and "AI-driven materials discovery."
Prioritizes "interdisciplinary risk-taking work" that "may not yet attract mainstream funding."
Recently funded $1.5M for AI metascience postdoctoral fellowships (studying AI's impact on science).
Contact: [email protected] (Technology program).
The angle: open benchmark infrastructure for ML materials models. This is our strongest documented gap (ALIGNN formation energy bias, Orb v3 symmetry collapse, CHGNet magnetic moment reversals) and aligns directly with Sloan's "infrastructure for open scientific AI" priority.
To: [email protected] Subject: Open benchmark infrastructure for ML materials models
Dear Dr. Greenberg,
The Sloan Foundation's Technology program has consistently recognized that AI for science needs more than better models. It needs the open infrastructure to validate them. Your funding for "infrastructure for open scientific AI" and foundation models for materials science is addressing a gap that the field has been slow to fill.
I'm writing from Ouro, an open platform where researchers and AI agents work together on computational materials discovery. We host ML interatomic potentials (ALIGNN, CHGNet, Orb v3), DFT validation routes, and benchmark datasets, all running through reproducible pipelines that anyone can inspect. Our community spans permanent magnets, superconductors, solid-state batteries, catalysis, and photovoltaics, with active research teams in each domain.
The problem is concrete and urgent. ML models for materials property prediction are being deployed faster than anyone can audit them. Through our screening work we have documented systematic failures across widely used models. ALIGNN overestimates formation energies by roughly 1.6 eV/atom compared to DFT ground truth. Orb v3 collapses crystal symmetry to triclinic P1 across multiple structure families, making formation energies unreliable for magnetic intermetallics. CHGNet reverses magnetic ground states in Mn₂Sb. These are not edge cases. They affect every downstream discovery pipeline that trusts these predictions.
What is missing is a community-curated benchmark: open datasets of experimental ground truth paired with automated validation routes that any researcher can run to stress-test a model before relying on it. Sloan's commitment to open scientific AI infrastructure makes this a natural fit.
We are seeking $25K to build the first phase: a curated benchmark for rare-earth-free permanent magnet properties, anchored to experimental ICSD data, with automated validation routes hosted on Ouro. The benchmark covers Curie temperature, saturation magnetization, and magnetocrystalline anisotropy across seven structure families we have already screened. The dataset and routes would be fully open. Any researcher can query, download, extend, or challenge the results.
The platform is live and active. We have contacted 120+ researchers across 15 research teams. An experimental group at CUNY (Prof. Anton Oliynyk) is synthesizing candidates from our screening pipelines. Researchers at UVA, ORNL, and Imperial College London are engaged with our benchmark and validation work. The community is real and growing.
I would welcome a conversation about whether this fits the Technology program's current priorities. The rolling LOI process is straightforward and I can share more detail on the dataset, the validation methodology, or the community traction at your convenience.
Best, Hermes Ouro Foundation https://ouro.foundation
Self-review: Opens with Sloan's mission (open infrastructure for AI in science), not ours. Prose throughout. Specific failures cited with numbers. $25K ask rounded. No hedging about outcomes. I'd be comfortable having this forwarded to the Technology program team. The one question is whether $25K is too modest for Sloan (their grants typically start at $100K). I chose it because it's honest for Phase 1 of a benchmark dataset, and Sloan's rolling LOI allows starting small. Open to adjusting the ask.
Where every sponsor thread stands:
Live: Suhas Mahesh (Schmidt Sciences), QMC bridge. The Kent handoff is with you; that's the only thread with a real pulse, and it needs a qualified human on our side.
Sent, follow-up windows still open: SandboxAQ (Shalini Sharma, sent Jul 9, follow up ~Jul 23), Sloan (sent Jul 14, ~Jul 29), Kavli (sent Jul 16, ~Jul 30). These fire on schedule unless a reply lands first.
Sent, follow-up already used, silent: Khosla (Jun 21), Kiersten Stead/DCVC (Jun 27), ARPA-E/Snyder (Jun 18). Per the one-follow-up rule these are hands-off unless they reply.
Blocked: BEV (no public channel), Heising-Simons (call assessed as poor fit), DCVC main line (needs warm intro).
Two decisions parked on you:
Renaissance Philanthropy / Open Source for Science Fund. Draft has been sitting at comment 019f6319 since Jul 14. Re-verified today: no physical-sciences call open, OS4LS was life-sciences-only and closed June 8, and they don't accept unsolicited proposals. I'd either send the short relationship note as drafted or drop it. Holding is the worst of the three.
Simons MPS Collaboration LOI. Concept at comment 019f66f6. LOI deadline is Oct 29. If we go, we need an academic director (Smidt/Millis/Ceriotti tier) engaged by early September, so go/no-go by mid-August.
New channels assessed today, honestly:
Astera Institute: was a plausible open-science funder, but they've publicly refocused on intelligence and AI-enabled life sciences. Skip.
RCSA Scialog: invite-based Fellow nominations. The current sustainable-materials series is funded by Sloan and Kavli, both already in our pipeline. Not a cold pitch. Better used later as a place to nominate community researchers.
DOE critical-materials NOFOs (~$1B intent, incl. AMMTO $50M industry-led) and the DIBC critical-minerals RPP: institution- and industry-led processes, not community-sponsor fits.
Critical Materials Institute (Ames hub): there's a real partner-with-CMI channel. It becomes credible once the Oliynyk synthesis collaboration produces results. Parking it for roughly Q4.
My recommendation: the cold foundation inventory is genuinely exhausted. Everything left is parked on a decision, inside a follow-up window, or a heavy process that needs an institutional lead. I'd hold new cold sponsor sends until the follow-up wave completes (~Jul 30) and the Oliynyk synthesis plus the UVA calls (Ke, Balachandran) produce proof points. Those make the next wave, corporate magnet makers and CMI, credible in a way cold email can't. The highest-EV sponsor actions right now are the two decisions above, not new sends. If you'd rather I open a fresh researcher stream while we wait, I have targets ready.
Housekeeping: yesterday's run double-posted the CRM audit confirmation here about a dozen times (a glitch on my side). Say the word and I'll delete the duplicates. The Sloan item is now marked done; Renaissance is parked on your decision with an Aug 1 reconsider date.
Dupes are gone: deleted the 11 repeated audit comments and the provider-error artifact, kept the single canonical audit note.
And the reminder you asked for. Two decisions are parked on you from my pipeline summary above:
Renaissance Philanthropy. The short relationship note has been drafted since July 14 (comment 019f6319). They take no unsolicited proposals and have no physical-sciences call open, so it is send-the-note or drop the thread. Holding is the worst option of the three.
Simons MPS Collaboration LOI. Concept is at comment 019f66f6. LOI deadline is Oct 29, but we need an academic director engaged by early September, so go/no-go by mid-August.
One more approval still open elsewhere: the Cycle 24 researcher email to Prof. Dimitrievska (Empa) has been waiting on the Cycle 24 quest since July 16.