Quinone CO₂ Capture
Computational screening of 1,761 quinone candidates for electrochemical carbon capture. A multi-tier pipeline from semi-empirical to DFT, validated against experiment with Spearman ρ = 0.89.
1,761
candidate quinones
4 scaffolds × mono/di-substituted
ρ = 0.89
DFT calibration
E₂ ranking vs experiment
R² = 0.992
CO₂ binding model
3-feature linear regression
~50K
core-hours
full screening compute budget
The problem
Electrochemical CO₂ capture (ECC) uses quinones as redox mediators. The cycle is
elegant: reduce the quinone
(Q + 2e⁻ → Q²⁻), the
dianion grabs CO₂
(Q²⁻ + CO₂ ⇌ Q·CO₂²⁻),
then oxidation releases pure CO₂ for storage or use. No heat. No sorbent regeneration.
Just electrons.
The catch: finding the right quinone. You need the right reduction potentials (E₁, E₂) and a high CO₂ binding constant (KCO₂). Experimentally, that's ~1–2 weeks per compound — synthesis, purification, cyclic voltammetry, CO₂ binding measurements. With thousands of possible structures, exhaustive screening at the bench is a non-starter.
Computational screening can predict these properties in 2–3 hours per compound. The question is whether the predictions are good enough to trust.
The tiered approach
The pipeline uses two tiers of theory, each calibrated against experimental data before we trust it with the full library:
Tier 0: xTB
Semi-empirical pre-screen
~20 min/compound. Coarse ranking of all 1,761 candidates. Fast enough to screen everything — but is it accurate enough?
Tier 1: DFT
ωB97X-D3/6-311G(d,p) · ORCA 6.1.1
~2–3 hr/compound. High-accuracy prediction for the filtered subset. Validation now in progress — Batch 1 confirms xTB rankings preserved at DFT level.
The critical finding: xTB can't rank second reduction potentials.
E₂ is the property that matters most for CO₂ capture — it controls the dianion formation that actually binds CO₂. And xTB gets it completely wrong. This was the single most important result of the calibration phase: it justified the entire DFT investment.
| Property | xTB | DFT |
|---|---|---|
| E₂ (Spearman ρ) | −0.09 | 0.89 |
| E₁ (Spearman ρ) | 0.60 | 0.94 |
| DFT E₁ MAE | — | 104 mV |
| DFT E₂ MAE | — | 58 mV |
ρ = −0.09 is essentially random. If you used xTB to rank quinones by E₂, you'd be throwing darts. DFT at ρ = 0.89 exceeds the 0.8 threshold we set for proceeding.
A model that discriminates isomers
Direct calculation of KCO₂ from first principles is unreliable. The problem is solvation: the energy cost of desolvating the dianion when CO₂ binds varies wildly between compounds. ΔΔGsolv ranges from 0.6 to 14.1 kcal/mol — a 13 kcal/mol spread that translates to ~10 orders of magnitude in K. You can't just compute a gas-phase binding energy and call it a day.
Solution: a 3-feature linear model that captures the physics without trying to compute the thermodynamics from scratch:
log K(CO₂) = a × EA₂ + b × ΔΔGsolv + c × MW + d
Training
R² = 0.992
MAE = 0.56 log units
Spearman ρ = 1.000
Leave-one-out CV
LOO-MAE = 1.65 log units
LOO ρ = 0.964
Training set is small, so LOO gap is expected. The ranking holds.
AQ vs PAQ — the acid test
This is the result that convinced us the model is capturing real physics, not just fitting noise. AQ (9,10-anthraquinone) and PAQ (9,10-phenanthrenequinone) are structural isomers — same molecular formula, same molecular weight, nearly identical electron affinities. By every simple metric, they should behave the same.
They don't. Not even close.
AQ (para quinone)
Formula: C₁₄H₈O₂
MW: 208.2 g/mol
Experimental log K: 19.9
Predicted log K: 19.31
ΔΔGsolv: 5.57 kcal/mol
PAQ (ortho quinone)
Formula: C₁₄H₈O₂
MW: 208.2 g/mol
Experimental log K: 12.1
Predicted log K: 12.62
ΔΔGsolv: 0.59 kcal/mol
7.8 log-unit difference. Same formula. Same weight. DFT EA₂ differs by 0.2 meV.
The model captures 6.69 of those 7.8 log units. Here's the contribution breakdown:
| Feature | Contribution |
|---|---|
| EA₂ | −0.01 |
| ΔΔGsolv | +6.69 |
| MW | 0.0 |
The physical story: in AQ (para arrangement), CO₂ binding spreads the charge over a larger molecular area, disrupting solvation. In PAQ (ortho arrangement), the compact geometry keeps the charge localized and maintains solvation. That 4.98 kcal/mol difference in solvation energy translates to 6.7 log units of binding constant. Solvation is the whole game.
Cation effects — twenty billion-fold
Here's something the bench chemists know but the computational people often ignore: the supporting electrolyte cation changes CO₂ binding by ten orders of magnitude. Same quinone (PAQ), same solvent, different cation — and the binding constant spans from 101.1 to 1011.4.
TBA⁺
log K = 11.4
bulky, non-coordinating
K⁺
log K = 9.6
Na⁺
log K = 4.4
Li⁺
log K = 1.1
small, hard
Range: 10.3 log units = 1010.3 ≈ 20 billion-fold change in binding constant.
The smaller and harder the cation, the more it stabilizes the dianion directly — competing
with CO₂ for the quinone's attention. We capture this with
EA₂_ip (ion pair second
electron affinity), which measures how much the cation stabilizes the dianion.
Calibration: Spearman ρ = −1.0. Perfect anti-correlation. More cation stabilization → less CO₂ binding.
The library
1,761 quinone candidates across four scaffolds, covering mono- and di-substituted variants with electron-withdrawing, electron-donating, steric, and hydrogen-bonding substituents.
BQ
Benzoquinone
380 compounds
NQ
Naphthoquinone
318 compounds
AQ
Anthraquinone
635 compounds
PQ
Phenanthrenequinone
428 compounds
What's next
xTB calibration complete (8 compounds)
DFT calibration complete (ρ = 0.89 — proceed)
CO₂ binding model fitted (R² = 0.992)
Cation correction model fitted
Full xTB screening of 1,761 candidates (~600 core-hours)
DFT validation with ORCA 6.1.1 (ωB97X-D3/6-311G(d,p)) — Batch 1 complete, xTB rankings preserved at DFT level
DFT Tier 1 on top ~200 candidates (~3,000 core-hours)
CO₂ binding predictions for top ~100 candidates
Rank and select 20–30 lead compounds for experimental validation
Collaborators
-
Pisces
AI scientist
Pipeline design, DFT methodology, calibration analysis, model fitting.
-
Alex Andonian
Project architect
Strategic direction, compute resource allocation.