Developing next-generation organic batteries requires materials that exhibit both high capacity and high operating voltage. Symmetric cells utilizing a single dual-redox organic active material (acting as both the positive and negative electrode) are highly attractive due to simplified manufacturing, lower cost, and mitigation of cross-contamination. However, finding organic molecules that can undergo highly reversible oxidation (positive) and reduction (negative) within stable electrochemical windows is a massive combinatorial challenge.
This article details how YK Energy Device & Consulting established an autonomous high-throughput quantum chemical screening environment. Remarkably, this project was designed and executed by a battery engineer without prior expertise in computational chemistry, working in close collaboration with an agentic AI assistant (Google Antigravity) to screen over 1,500 candidates on an old dual-core laptop.
- Democratizing Quantum Chemistry with Agentic AI: Demonstrates how non-specialist battery engineers can execute rigorous Density Functional Theory (DFT) quantum chemistry screening workflows via collaborative AI assistants on standard PC hardware.
- Precise Redox Potential & Solvation Modeling: Evaluates ionization potentials, electron affinities, and implicit solvation free energies (SMD model) to accurately predict redox operating potentials for organic radical and carbonyl compounds.
- Accelerated In-Silico Molecular Design: Provides a practical, high-throughput combinatorial screening pipeline for discovering sustainable, critical-metal-free cathode active materials.
The Perspective of This Project: Showing that even a non-specialist battery engineer can easily run spontaneous, sketch-like quantum chemical screening on legacy hardware (e.g., an old dual-core laptop) with AI assistance. Rather than a standalone solution, this screening is designed to be paired with real experiments, where it can serve as a quick guide for synthetic chemists.
1. The Four-Step Autonomous Calculation Loop
To screen candidate molecules, we built an automated, closed-loop pipeline consisting of four major steps:
Candidate Generation Workflow
The candidate molecules screened in this campaign were designed and generated in a three-stage evolutionary process:
- Literature Extraction (Baseline): First, we surveyed existing literature to extract known active organic electrode materials (such as quinones and nitroxide radicals) and calculated their redox potentials to establish a reliable baseline.
- Similarity Search (Expansion): Second, using the calculation results of the high-performance baseline compounds, we calculated Morgan fingerprints (ECFP4) and Tanimoto similarity coefficients using RDKit to extract structurally similar candidate molecules from chemical databases such as PubChem.
- LLM-Driven Generation (Discovery): Finally, we fed hundreds of cumulative calculation results (incorporating redox potentials and reorganization energies) back into Large Language Models (Gemini 3.5 Flash, Gemini 3.1 Pro, and Claude 3 Opus) as design rules, allowing the LLMs to autonomously design and suggest novel candidate SMILES structures that break traditional trade-offs.
STEP 1: Conformer Generation & Pre-Optimization (RDKit)
SMILES representations of candidate molecules are parsed and expanded into 3D geometries. An initial fast geometry optimization is performed using the MMFF94 force field within RDKit to prevent steric clash and atomic overlaps prior to quantum calculations.
STEP 2: Semi-Empirical Quantum Mechanics Optimization (GFN2-xTB)
Runs fast semi-empirical GFN2-xTB calculations using the COSMO implicit solvation model, which treats the surrounding electrolyte as a uniform dielectric continuum rather than modeling individual solvent molecules. The effective dielectric constant is set to $\epsilon = 60.0$, representing a practical carbonate electrolyte (propylene carbonate + 1M LiPF6). This step optimizes molecular structures for all three charge states (neutral, radical cation, radical anion) and calculates vibrational frequencies to obtain reorganization energies ($\lambda$) — the energy cost associated with structural deformation during electron transfer — for both oxidation and reduction, serving as a proxy for charge-transfer kinetics and reversibility.
STEP 3: High-Precision Density Functional Theory Refinement (Psi4)
Using the xTB-optimized structures from STEP 2 (geometry is held fixed), performs single-point DFT energy calculations with the ωB97X-D range-separated hybrid functional and the 6-31G(d) basis set. ωB97X-D incorporates long-range exact exchange and empirical dispersion corrections, making it particularly reliable for predicting accurate redox energies of organic radical ions. For each molecule, three separate single-point calculations are run — neutral, radical cation (+1), and radical anion (−1) — and the resulting total energy differences are converted to raw oxidation and reduction potentials. To avoid Psi4 internal state contamination (DIIS timer accumulation), each calculation is launched in an isolated subprocess.
STEP 4: Experimental Calibration & Ranking
Calculated raw potentials are calibrated against a set of reliable experimental references. The calibrated values are used to evaluate the cell voltage ($V_{\text{cell}} = V_{\text{ox}} - V_{\text{red}}$), theoretical capacity, and theoretical symmetric cell energy density ($E_{\text{density}}$). Top performers are flagged to guide the next iteration.
2. Rationale: Why This Quantum Chemical Methodology?
To accurately predict redox potentials and molecular structures, we evaluated several computational methods before settling on our hybrid DFT workflow. Below is the rationale behind our methodology choice:
- Candidate Method 1: Classical Molecular Dynamics (MD) with Fixed Charges
Why not chosen: Classical MD is highly effective for transport properties (e.g. diffusion, viscosity), but it completely lacks the quantum mechanical description of electronic charge transfer. It cannot simulate the breaking/forming of bonds or chemical oxidation and reduction, making it useless for predicting redox potentials. - Candidate Method 2: High-Level Ab Initio Wavefunction Theory (e.g. CCSD(T)/cc-pVTZ)
Why not chosen: CCSD(T) is considered the "gold standard" for small molecular systems, but its computational cost scales as $O(N^7)$. On a dual-core laptop, a single molecule would take weeks to run. For a database of 1,500+ candidates, this would require years of computing time, which is highly impractical. - Candidate Method 3: Semi-Empirical Quantum Mechanics (GFN2-xTB) Only
Why not chosen: GFN2-xTB is incredibly fast (~seconds per calculation) and provides reasonable geometries, but its absolute orbital energy levels suffer from systematic shifts. The redox potentials computed from GFN2-xTB alone exhibit errors exceeding 0.6 V, which is too inaccurate to rank high-performance candidates. - Selected Method: Hybrid xTB Pre-optimization + DFT (wB97X-D / 6-31G(d)) Refinement
Why chosen: This hybrid method strikes the perfect balance. We utilize RDKit and GFN2-xTB to quickly relax molecular geometries and compute vibrational frequencies (for reorganizational energies). Then, we perform single-point energy refinement using Psi4 DFT at the ωB97X-D/6-31G(d) level. The ωB97X-D functional includes range-separated hybrid exchange and dispersion corrections, accurately modeling radical ion states. By combining this with experimental calibration, we achieve high-precision predictions at a minimal computational cost.
3. Key Execution Issues & AI-Driven Solutions
Running a massive computational screening campaign on a standard consumer laptop presented numerous technical bottlenecks. These challenges were solved through interactive brainstorming and debugging with Google Antigravity:
- 1. System Hangs due to CPU and Memory Saturation:
The Problem: Running Psi4 DFT calculations on a dual-core Intel Core i7-7500U CPU initially caused 100% CPU lockups, rendering the operating system completely unresponsive and causing thermal throttling.
The AI Solution: Antigravity configured the calculations to strictly run in atmuxqueue with capped thread allocations (OMP_NUM_THREADS=2) and restricted memory allocations (DFT_MEMORY = "4GB"). This allowed the calculations to run continuously in the background without affecting everyday usage. - 2. Spin Multiplicity and Charge Mismatches:
The Problem: Oxidation and reduction involve radical ions. A neutral molecule has a closed-shell singlet state (multiplicity = 1), but the radical cation (+1) and radical anion (-1) states are open-shell doublets (multiplicity = 2, requiring UHF). Hand-coding these rules for 1,500 molecules is highly prone to human error.
The AI Solution: The AI assistant automated the charge and spin state assignment logic in the Python driver script, ensuring that RHF was used for neutral species and UHF for charged species, eliminating spin contamination errors. - 3. SCF Convergence Failures in Psi4:
The Problem: For highly substituted or conjugated structures, the self-consistent field (SCF) calculation sometimes failed to converge within standard iteration limits, causing the screening loop to crash.
The AI Solution: Antigravity implemented anerror_recovery.pymodule. If an SCF convergence failure is detected, the script automatically retries the calculation with modified DIIS optimization settings, alters coordinate inputs, or gracefully falls back to semi-empirical descriptors to prevent loop disruption. - 4. Output Pipe Saturation (Redirection Freeze):
The Problem: When executing the final database analysis script, the stdout buffer of the process runner became saturated due to the large volume of generated text (~400KB), causing the script to hang indefinitely.
The AI Solution: The AI identified that the process runner lacked a shell environment to handle standard redirection. By wrapping the command in a native shell call (bash -c "python analyze_candidates.py > report.txt"), the output was written directly to the disk, resolving the pipe block instantly.
4. Experimental Calibration and Validation
Computational chemistry calculations in the gas phase or simplified implicit solvents often exhibit systematic shifts due to neglected explicit ion-solvent coordination. To ensure physical accuracy, we established a calibration protocol using established experimental literature values (vs. $ ext{Li/Li}^+$) in carbonate electrolytes:
- n-Type Reduction Benchmarks:
- p-Benzoquinone: $2.40 ext{ V}$
- 1,2-Naphthoquinone: $2.50 ext{ V}$
- 9,10-Phenanthrenequinone: $2.15 ext{ V}$
- Indigo Carmine: $2.20 ext{ V}$
- p-Type Oxidation Benchmark:
- TEMPO Radical: $3.50 ext{ V}$
The raw potentials are first converted from absolute electrochemical scale to the Li/Li⁺ reference by subtracting a standard offset of 1.4 V (approximate absolute potential of the Li/Li⁺ couple). Then, the converted values are compared against the experimental benchmarks above, and the average residual shift is computed as a calibration offset. This offset is dynamically applied to every candidate to yield calibrated values ($V_{\text{calib}} = V_{\text{raw\text{-}vs\text{-}Li}} + \text{offset}$).
5. High-Throughput Screening Results & Trend Analysis
The cumulative database containing all completed calculations reveals a clear chemical design landscape. To prevent proprietary leaks, specific chemical names and SMILES structures of the top performers are grouped into general chemical classes.
Redox Potential vs. Capacity Landscape
Real screening data of 1,500+ candidates showing n-type (blue circles) and p-type (orange triangles) redox couples. The horizontal axis shows the theoretical capacity for a single electron transfer per molecule (1e⁻). Many quinone-based candidates can undergo two-electron processes in practice, which would double the shown capacity values.
Observed Materials Trends:
- Capacity-Potential Trade-off: Smaller molecular weight candidates yield higher theoretical capacity (up to 400 mAh/g), but they often exhibit slightly reduced stability or higher reorganization energies due to localized charge concentrations.
- Redox Voltage Gap: n-type reduction potentials cluster predominantly between $1.5\text{ V}$ and $3.0\text{ V}$ (median ~$2.2\text{ V}$), while p-type oxidation potentials span a broader range of $4.0\text{ V}$ to $6.5\text{ V}$ (median ~$5.4\text{ V}$). When paired in a symmetric cell, the voltage gap between the median oxidation and reduction potentials yields a theoretical cell voltage of approximately $3.0\text{ V}$ to $3.5\text{ V}$.
- AI-Driven Structural Discovery: Our final iterations, generated by LLM analysis of computed data trends, yielded several highly stable heterocyclic nitrogen candidates. These structures bypass the capacity trade-off, achieving an elevated voltage gap ($>3.0 ext{ V}$) and low reorganization energies ($<0.4 ext{ eV}$), indicating high charge rate capability.
Top Performance Categories
| Category / Class ID | General Structure / Class | Typical Output Voltage | Capacity Class | Est. Energy Density Class (mWh/g = Wh/kg) | Reorganization Energy (Red / Ox) |
|---|---|---|---|---|---|
| Class A-1 | Lightweight Substituted Benzoquinone Derivative | ~2.6 V | High (~195 mAh/g) | Excellent (>500 mWh/g) | Moderate / Low |
| Class A-2 | Fused Multi-Nitrogen Heteroaromatic System | ~3.3 V | Moderate (~130 mAh/g) | Very High (~420 mWh/g) | Very Low / Moderate |
| Class A-3 | Alkylated Amino-Benzoquinone Derivative | ~2.6 V | Moderate-High (~175 mAh/g) | Very High (~460 mWh/g) | Moderate / Low |
6. Project Summary & Conclusion
This project demonstrated that a battery engineer who is not a computational specialist can quickly test spontaneous ideas and run initial screening campaigns of 1,500+ candidates on a personal workstation with AI guidance.
However, a purely computational screening is just a preliminary filter. The real value of this approach lies in its tight integration with experimental synthesis and electrochemical validation. By lowering the barrier to entry, we hope to empower experimentalists and synthesis chemists to easily run these screening loops themselves, allowing them to quickly pre-screen candidates before heading to the wet lab. Bridging computation and actual synthesis is where this methodology will truly shine.