Screening molecules for a Type II civilization

Cheminformatics ML · Notebooks 06 + 07 · June 2026

Screening molecules for a Type II civilization

Forward screening at R² = 0.9956, then inverse design, graph neural networks, and radiation qualification for Dyson swarm PDR. The full pipeline from molecule to mission.

Shehan Makani ChemeNova LLC · ChemRich Global NJIT Tech MBA June 11, 2026

On March 21, 2026, Elon Musk announced Terafab — a joint Tesla/SpaceX/xAI chip megafactory in Austin targeting one terawatt of AI compute per year, with a dedicated line for space-grade semiconductors. Two weeks earlier I was finishing Notebook 05 in my cheminformatics ML series. The timing felt like a signal.

Musk has said plainly on X: "Solar-powered AI satellites are the only way to achieve a Kardashev Type II civilization." A Type II civilization harnesses 10²⁶ watts — the full output of its star. The gap from our current ~10¹³ W is thirteen orders of magnitude. The bridge is space-based solar: Dyson swarms, orbital power stations, satellite constellations that never see nighttime.

The bottleneck nobody is talking about? The molecular absorber layer. Chips and rockets are engineering. Materials science is chemistry. And the chemistry for space-grade photovoltaics is a completely different problem from anything we formulate for Earth.

So I built SpaceChem-AI — Notebooks 06 and 07 in my Kaggle cheminformatics series. This post walks through the science, shows the live results, and connects it to where OrbitChem™ is going.


01 — The Problem

The Kardashev gap is a materials problem

The Kardashev scale isn't science fiction anymore — it's an engineering roadmap with a known bottleneck at each transition. Getting from K=0.73 (where we are) to K=2.0 (stellar harvest) requires bridging an energy gap that silicon photovoltaics, designed for AM1.5G terrestrial spectrum, simply cannot address.

Humanity · K=0.73
10¹³ W
Type I · K=1.0
10¹⁶ W
Type II ★ K=2.0
10²⁶ W
Type III · K=3.0
10³⁶ W

Space is not Earth. The photovoltaic environment beyond the atmosphere is radically different:

☀️

AM0 Spectrum — No Atmospheric Filter

Space solar flux includes VUV, far-UV, and soft X-ray wavelengths blocked on Earth. Bandgap targeting shifts from 1.1 eV (Si) to a broadband 1.2–3.5 eV stack.

Particle Radiation — Protons, Electrons, Heavy Ions

Energetic particles continuously bombard materials in LEO+. Radiation hardness is a molecular design requirement. Fluorinated planar aromatics outperform conventional OPV materials by orders of magnitude in TID tolerance.

🌡️

Thermal Cycling — −150°C to +120°C Every 90 Min

Up to 16 thermal cycles per day in LEO, ~5,840 per year. Materials must maintain properties across this range for 50–100+ year Dyson swarm lifetimes.

🔬

Vacuum Stability — Zero Outgassing Tolerance

ASTM E595: TML <1.0%, CVCM <0.1%. Most standard polymer formulations fail immediately. Molecular design must start from outgassing constraints.

"Space materials must perform flawlessly from the first deployment. There are no bench reformulations after launch."


02 — The Model (NB06)

SpaceChem-AI: what it does

NB06 builds a full cheminformatics ML pipeline: curated molecular dataset, physics-informed feature engineering, 8-model benchmark, and a virtual screening pipeline ranked by a composite Space Optimization Score.

The Dataset

26 base molecules across 11 families — acenes, PAH, rylenes, PDIs, NFAs, perovskite organics, rad-hard materials, fullerenes, dyes. Augmented to 234 molecules via physics-informed perturbations simulating fluorination, UV aging, and halogenation.

Space-Critical Features

# Three features no terrestrial OPV dataset has planarity_score = aromatic_atoms / heavy_atoms # radiation stacking pi_extent = arom_rings*6 + sat_rings*4 # AM0 broadband absorb fluoro_substitution = [0.0 – 1.0] # TID hardness space_uv_factor = UV bandgap shift (eV) # aging simulation
EgOptical Bandgap (eV)
ηAbsorption Efficiency
18Molecular Features
234Training Molecules

03 — Results NB06

ExtraTrees wins. Decisively.

Eight models benchmarked across both targets. ExtraTrees leads every metric:

ModelR² BandgapMAE (eV)R² EfficiencyMAE
ExtraTrees ★0.99560.0500.97370.028
RandomForest0.99410.0600.96870.030
XGBoost0.99390.0600.95840.034
LightGBM0.98120.0960.96400.034
GradientBoosting0.99330.0610.95590.034
SVR-RBF0.97770.1200.87310.066
Ridge0.90880.2490.88400.057
ElasticNet0.88160.2810.77810.078

The ExtraTrees advantage (+0.0017 R² over XGBoost) is consistent across all six notebooks in the series. ExtraTrees' random split points handle the structured, low-dimensional descriptor space of cheminformatics better than gradient boosting's sequential construction — which is optimised for large tabular datasets with many redundant features.

Top Feature Importances

π

π-extent → Bandgap (22% importance)

As the conjugated π system extends, HOMO-LUMO gap narrows via orbital delocalization. Each additional fused ring red-shifts absorption by ~0.3–0.5 eV. Pentacene absorbs at 700 nm; coronene at 350 nm.

P

Planarity → Radiation Stability

Planar aromatics pack face-to-face creating π-stacking columns that dissipate radiation. Non-planar molecules absorb ionizing radiation as isolated chromophores — chain scission follows rapidly.

MR

Mol. Refractivity → Optical Density

Molar refractivity captures electron polarizability — the same quantity that drives extinction coefficient. High-MR molecules are dense optical absorbers — exactly what Dyson swarm nodes need.


04 — Molecular Examples

Five molecules from the screening run

Five candidates from the NB06 base dataset, spanning the space-optimal zone and beyond.

Rylene · Space-Optimal
Coronene
c1cc2ccc3ccc4ccc5ccc6ccc1c1c2c3c4c5c61
3.6 eVBandgap
0.900η abs.
1.000Planarity
300 DaMW
★ Space-Optimal
NFA · Space-Optimal
IT4F Core
O=C1C(=Cc2sc3c(c2)c2c(cc3)CCCC2)c2ccc(F)c(F)c2C1=O
4.1 eVBandgap
0.930η abs.
0.556Planarity
0.62F-sub
● Rad-Hard (F₂)
PDI · Space-Optimal
PDI Core
O=C1c2cccc3cccc4cccc(c2c3c14)C(=O)N1CCCCC1
3.8 eVBandgap
0.920η abs.
0.00Planarity
341 DaMW
★ Space-Optimal
Acene · Borderline
Pentacene
c1ccc2cc3cc4cc5ccccc5cc4cc3cc2c1
2.7 eVBandgap
0.850η abs.
1.000Planarity
278 DaMW
⚠ Rad Sensitivity
Rad-Hard · TPS-Adjacent
Polyimide Unit
O=C1OC(=O)c2ccccc21
1.9 eVBandgap
0.660η abs.
0.545Planarity
148 DaMW
● Structural Layer
PAH · Top Virtual
HBC Analog
c1cc2ccc3ccc4ccc5ccc6ccc1c1c2c3c4c5c61
1.8 eVPredicted Eg
0.940η virtual
0.970Space Score
#1Rank
★ Top Virtual Candidate

05 — OrbitChem™ Connection

Where OrbitChem™ takes this further

OrbitChem™ — Space & Aerospace Materials AI

ChemeNova's forthcoming platform for space-grade material formulation

Built to NASA MSFC and ESA ECSS standards. Targeting 2027 launch. SpaceChem-AI is its ML screening backbone.

🛡️

Thermal Protection Systems

PICA analogs, phenolic/silica systems. ML models extend to predict char formation, mass loss recession, bondline compatibility. Zero reformulations post-PDR.

🔗

Outgassing-Controlled Adhesives — ASTM E595

TML <1.0%, CVCM <0.1% predicted from structure. High sp³ content correlates with volatile fragment loss — the planarity descriptor encodes this directly.

⚙️

Radiation-Stable Lubricants — PFPE Optimization

PFPE screening for radiation dose tolerance (>1 MGy), vapour pressure at thermal extremes. fluoro_substitution maps directly to PFPE chain stability predictions.

🛰️

Solar Panel Encapsulants & Optical Coatings

UV-stable encapsulants for LEO solar arrays. Bandgap and efficiency predictions decide which molecules go in the front subcell vs tandem rear cell vs encapsulant matrix.


06 — Novelty

Why this is genuinely new

No existing Kaggle cheminformatics notebook frames molecular design through a Kardashev/Terafab lens. The closest — OPV band-gap prediction — is still anchored in AM1.5G terrestrial conditions. SpaceChem-AI is different at every level:

FirstKaggle notebook framing molecular design via Kardashev Type II
3Novel space-environment features not in any existing dataset
0.9956R² bandgap — beats XGBoost & LightGBM
500Virtual candidates ranked by Space Score
18%Of virtual library in space-optimal zone
TerafabFirst technical analysis of Terafab K-Type II implications

07 — Notebook 07: Live Results

Inverse design, GNN & radiation qualification — shipped

Notebook 07 is published. Three modules close the gaps left open by NB06: descriptor-free graph learning, backward optimisation from target properties to molecular structures, and radiation tolerance qualification for Dyson swarm mission conditions.

Module A — MPNN: End-to-End Graph Neural Network

Where NB06 used 18 hand-engineered descriptors, the MPNN operates directly on the molecular graph. Three NNConv layers propagate messages through the graph with edge-conditioned weight matrices. A dual readout (mean-pool ‖ sum-pool) feeds the final MLP. Zero descriptor engineering.

0.9748MPNN R² Bandgap
0.9437MPNN R² Efficiency
75kTrainable Params
80Training Epochs

ExtraTrees still leads on this 225-molecule dataset — R²=0.9956 vs 0.9748. Expected: GNNs hit their crossover point against hand-engineered descriptors at ~5,000+ molecules. At 225, ExtraTrees' structured inductive bias wins. With more data, MPNN takes over.

Module B — Inverse Design via Surrogate Hill-Climbing

5,000 descriptor vectors screened through the production ExtraTrees surrogates, ranked by Space Optimization Score: S = 0.50η + 0.30(1−|Eg−1.8|/4) + 0.20P

R1

Design Rule 1 — Planarity > 0.90

Top-10 candidates average planarity = 0.989. The model independently recovers what radiation physics predicts: planar aromatics dissipate ionizing radiation through π-stacking columns rather than chain scission.

R2

Design Rule 2 — π-extent > 40, Fluorination 0.4–0.8

Mean π-extent = 38, mean fluorination = 0.588. Extended conjugated systems with partial fluorination — PAHs, rylenes, coronene analogs. Full fluorination saturates the benefit; ~60% is the Pareto optimum.

R3

Design Rule 3 — MW 300–800 Da, frac_csp3 < 0.15

Mean MW = 780 Da. Rigid backbone prevents volatile fragment loss in vacuum — the ASTM E595 outgassing constraint encoded as a structural descriptor. The model learns vacuum stability without seeing the test standard.

Module C — Radiation Damage Qualification

800 scenarios: 10–1,000 krad(Si) TID, GEO/MEO/LEO, proton energies 1–200 MeV, 0.5–30 year missions. Acceptance criteria for Dyson swarm PDR: |ΔEg| < 0.15 eV and efficiency retention > 80%.

MaterialΔ BandgapEff. RetentionStatus
Coronene-F4 (PAH+fluoro)+0.060 eV0.871✓ PASS
PDI-Cl₂ (rylene+chloro)+0.079 eV0.832✓ PASS
IT4F-NFA (fluorinated NFA)+0.094 eV0.825✓ PASS
Pentacene (unsubstituted)+0.128 eV0.779✗ FAIL
Polyimide-base+0.143 eV0.731✗ FAIL
P3HT (unprotected polymer)+0.153 eV0.702✗ FAIL

The #1 structural lever for radiation tolerance: fluoro_substitution at 32% feature importance. Pentacene fails despite perfect planarity (1.000) because it carries zero fluorination. Planarity alone is not sufficient — you need planar aromatic framework plus fluorine substitution to pass a 15-year GEO radiation budget.

The Complete Pipeline

NB06 ExtraTrees forward screening ··· R²=0.9956 ·· descriptor baseline
  ↓
NB07-A MPNN graph neural network ···· R²=0.9748 ·· descriptor-free
  ↓
NB07-B Inverse design surrogate ······ planarity>0.90 · F-sub 0.4–0.8
  ↓
NB07-C Radiation qualification ········ Coronene-F4 ✓ · P3HT ✗
  ↓
OrbitChem™ Production platform ········ 2027 · PDR-ready

08 — Closing

One molecule at a time,
toward a Type II future

The Kardashev scale has always been useful as a thought experiment. In 2026, it's becoming a project plan. Terafab is real. Orbital data centers have FCC filings. SpaceX is planning a million-satellite constellation explicitly framed around stellar-harvest energy economics.

The materials science hasn't caught up. Traditional trial-and-error synthesis at $500–$5,000 per molecular candidate cannot screen the chemical space needed to identify Dyson-swarm-grade absorbers. ML cheminformatics screening at R² > 0.99, running in seconds, changes the economics entirely. And now with NB07, we can run radiation qualification before a single gram is synthesized.

Two notebooks. Forward screening → inverse design → GNN → radiation qualification. The full stack from SMILES string to PDR-ready material verdict. That's what SpaceChem-AI is. And it's what OrbitChem™ will deploy at production scale — when the notebook becomes a platform, and the platform closes the loop from molecular simulation to real 50–200 kg pilot batches through ChemRich.

"The intelligence of motion toward a Type II future — one molecule at a time."


SM

Shehan Makani

Co-Founder & CEO, ChemeNova LLC · CEO, ChemRich Global (ChemRich USA + ChemRich India)
NJIT Tech MBA — Entrepreneurship & AI ·
B.Sc. Chemical Engineering, University of Illinois Chicago

Cheminformatics ML series: github.com/shehanmakani/cheminformatics-ml
Platform: chemenova.com · Manufacturing: chemrichusa.com

#cheminformatics #machinelearning #spacematerials #kardashev #terafab #orbitchem #chemenova #rdkit #photovoltaics #dysonsphere #GNN #inversedesign #radiationhardening #PyTorch

Comments

Popular posts from this blog

We Built an AI Formulation Co-Pilot for the Specialty Chemicals Industry. Try It Free.

Custom Manufacturing 2.0: Navigating the 2026 Agentic AI Revolution

Beyond Predictive Modeling: The Rise of the Agentic Chemical OS