Research portfolio · ChemeNova · October 2026

Fluids for heat and vacuum

Two research modules, run on ChemeNova's IntelliForm and OrbitChem platforms. One designs PFAS-free immersion cooling fluids for AI data centers. The other maps where space chemistry has a real, open gap, then models it from the molecule up.

Both projects are the same problem at different temperatures: predict how a fluid's viscosity, volatility and heat transfer change with temperature, then design the blend.
10immersion fluids screened against the OCP specification
12,842NASA outgassing tests analyzed
0.81AUC predicting test pass or fail on products never seen
×2.5worst miss of the physics model across five fluids, with no fitting
Project 1 · AI infrastructure

ThermalForm: PFAS-free immersion cooling fluids

AI racks now run hot enough that servers are submerged in dielectric fluid. The best-known fluids are fluorinated, and PFAS rules are closing in. ThermalForm, a new IntelliForm module, designs non-fluorinated blends that meet the Open Compute Project (OCP) immersion-fluid specification and cool better than the PAO benchmark.

What I built

  • Blend property models from published data sheets: viscosity (Refutas, Walther), flash point, thermal conductivity, heat capacity and permittivity.
  • Heat-sink physics for a 1 kW chip: laminar developing flow in plate fins, fin efficiency and pressure drop.
  • A 1 MW facility loop: plate heat exchanger, piping and pump energy.
  • Signal integrity: how a fluid shifts a 50 Ω trace's impedance.
  • Multi-objective optimization (NSGA-II) under OCP limits, with Monte Carlo on the uncertain inputs.
  • A browser fluid selector that reproduces the Python models to within 0.01%.

Findings that changed the design

  • The bottleneck is a dielectric number. OCP caps permittivity at 2.3 up to 40 GHz. Ester permittivity at those frequencies is unpublished, so it decides how much ester a blend can carry. The all-ester blend from round 2 passes in only about 5% of simulated cases.
  • Pumping is not negligible. At facility scale it is about 2% of IT power, roughly $20,000–34,000 per MW a year at New Jersey prices. Round 3 corrected round 2 on this.
  • The flow is laminar. Conductivity and heat capacity matter more than viscosity at the chip, which overturns the usual turbulent figure of merit.

Round 3 lead blends against the PAO benchmark

LeadRecipeChip-side ΔTPumping, $/MW·yrWorst-case DkFlash °CEdge
Aall-roundPAO 4 40% + Priolube 1976 35% + PAO 2 25%−0.40 °C−3,1002.28186Cooler in 97% of Monte Carlo draws
GbiodegradableRenewable isoparaffin 62% + Priolube 1976 38%−0.25 °C−5,5002.29170100% readily biodegradable
Ppremium flashPAO 4 81% + EF 7010 19%−0.28 °Cabout par2.22216Meets a 210 °C flash class

Benchmark: PAO 4 88% + PAO 2 12%. Model check: PAO 4's OCP figure of merit at 50 °C comes out at 45.4 against Evonik's published 46. Flash points are model values; prices are placeholders pending supplier quotes.

Run on IntelliForm v1.6

I added ThermalForm to IntelliForm as a new vertical, so the physics runs inside IntelliForm's NSGA-III Pareto engine and its TOPSIS selector. Without being told the earlier answer, IntelliForm found 35 Pareto blends. Its coolest fully biodegradable blend is renewable isoparaffin 62.9% + Priolube 1976 37.1%, which is Lead G to within a point.

  • For the 210 °C flash class, TOPSIS picked PAO 4 55% + Priolube 1976 38% + renewable isoparaffin 8%: 0.76 °C cooler than PAO.
  • Several picks ride exactly on the Dk 2.3 limit, so the final choice needs a safety margin until ester GHz permittivity is measured.
  • IntelliForm's QSAR ranks esters above hydrocarbons but rates every hydrocarbon near 75, agreeing with data sheets on only 5 of 9 stocks. Base oils are outside its training data.
  • IntelliForm's digital twin puts a 1-tonne pilot batch at 2.6 hours and 0.031 kg CO₂ per kg, a simple blend ChemRich USA can make.
Project 2 · Spacecraft materials

Space chemistry with OrbitChem

I scanned eight space-chemistry niches for a gap ChemeNova could own, re-ranked them for how crowded each story is, and went deep on the two that share one science: how molecules behave in vacuum.

Eight niches scored

Market pull, timing, fit, capital, qualification burden and competition. Lubricants and low-outgassing materials led at 35 and 34 out of 45.

Then re-ranked for crowding

PFAS-free is a crowded story; outgassing prediction is quiet. With crowding scored, outgassing moved to first (39 of 50) and lubricants to second (37).

One engine for both

Lubricant life and the outgassing test both come down to vapor pressure and molar mass, so one vacuum-materials engine, now part of OrbitChem, serves both.

The PFAS-free lubricant gap

Cold end: viscosity reaches 10,000 cSt. Hot end: evaporation removes 5 mg/cm² over 15 years. The best PFAS-free oil beats PFPE when hot and falls 16 °C short when cold. That defines the OrbitChem target: ≤ 10,000 cSt at −40 °C and ≤ 1×10⁻¹¹ Pa at 25 °C.

What 12,842 NASA tests say

Only 46% of tested materials pass (TML ≤ 1.0%, CVCM ≤ 0.10%). 37% of failures are near misses, and devolatilized versions pass 69% of the time against 45% for the rest.

A model that screens before the lab

Trained on what a buyer knows before testing (material class, usage, maker, cure and the product name) and always scored on product families it had never seen.

  • AUC 0.81 and 74% accuracy on unseen families; 0.74 on unseen brands.
  • The test itself agrees with repeat tests only 86% of the time, so that is the ceiling.
  • Sending only the model's top 30% to the lab lifts the pass rate from 46% to 79%.
  • The public data stops at 74% because it records trade names, not recipes. Composition is the missing input, and a formulator has it.

A physics model of the test itself

The ASTM E595 test cell is an effusion cell: vapor leaves through a 6.3 mm port toward a collector at 25 °C. From each component's vapor-pressure curve and molar mass, the simulator integrates 24 hours at 125 °C.

  • Mass loss lands within 2.5× of NASA's measurements across a more than 100,000-fold range of volatility, with nothing fitted.
  • Condensables match three of five fluids and run high on the other two, so the model errs on the safe side.
  • NASA's PFAS-free silahydrocarbon and the OrbitChem target oil are both predicted to pass with 3–11× margin.

Physics model against NASA measurements

Shaded band: within 2.5× of measured. Measured values are NASA medians. The polyphenyl ether uses Santovac 5 vapor-pressure data; NASA's entry is the related Convalex-10.

Design rule for a single low-volatility fluid · 125 °C · 24 h
P125 °C × √M < 2.3×10⁻⁵ Pa·(kg/mol)½ to pass CVCM ≤ 0.10%

For a 900 g/mol oil that means vapor pressure under about 2.4×10⁻⁵ Pa at 125 °C. Pennzane sits about 9× above the line, matching its mixed test record; PFPE sits about 5× below. This turns a three-day qualification test into a target OrbitChem can design against.

Built into OrbitChem

OrbitChem's app listed a standalone ASTM E595 outgassing screen as coming next. I built it: the simulator is ported to TypeScript and runs in OrbitChem's /outgassing page with presets, a component editor, a per-component breakdown, the design rule and the NASA validation table. It reproduces the Python results and builds cleanly.

Where it leads as a business

Outgassing pass engineering: screen a material before the three-day test, trace a failure to the volatile component, and fix it by devolatilization, post-cure or reformulation. Buyers are material suppliers seeking space qualification and NewSpace teams building with commercial materials. PFAS-free lubricants stay as the hard flagship application of the same engine, funded first through NASA SBIR and AFRL programs.

Methods and tools

How it was built

PythonNumPy · SciPypandasscikit-learnIntelliForm NSGA-III · TOPSISIntelliForm QSAR · digital twinOrbitChem (Next.js · TypeScript)pymoo NSGA-IIMonte CarloJavaScriptPlaywrightWalther · Refutas · FilippovHertz–Knudsen · Clausius–ClapeyronVDI plate heat exchangerHammerstad–Jensen microstripASTM E595 · OCP immersion spec

Every input carries its source and basis, and every model was checked against something published before it was trusted: Evonik's figure of merit, measured viscosities for EC-120, NASA bearing and vapor-pressure reports, and the NASA Goddard outgassing database. Where data was missing, the work says so and names the measurement that would close it.

From ChemeNova

OrbitChem: molecular design for space

The vacuum-volatility engine on this page is part of OrbitChem, ChemeNova's platform for designing space-grade fluids and materials. See what it does and how to work with us.