AI-assisted discovery of high-temperature dielectrics for energy storage

Electrostatic capacitors are vital for high-power applications across defense, aerospace, energy, and transportation due to their unparalleled power density. However, a key challenge is enhancing discharged energy density (U) at elevated temperatures to reduce size, weight, and cooling requirements. Today’s benchmark dielectric, biaxially oriented polypropylene (BOPP), offers low loss and large electronic band gap due to absence of π-stacking moieties, but suffers from low dielectric constant and poor mechanical stability at high temperature, leading to rapid degradation of U with increasing temperature. Commercial high-temperature alternatives (e.g., polyimide, PEEK, PEI, FPE) often sacrifice band gap and thus energy density. The research question is how to design polymer dielectrics that simultaneously achieve high glass-transition temperature (Tg), large band gap (Eg), and moderate-to-high dielectric constant (ε), yielding high U at high temperatures. The study frames materials search as an optimization problem over polymer chemical structures and introduces an AI-driven pipeline, polyVERSE, to generate synthetically accessible polymers from available monomers, predict key properties, and screen candidates. The purpose is to fill the performance gap for 85–200 °C operation where no commercial solution exists, reducing cooling system needs and enabling broader deployment. The significance lies in combining AI with reaction-aware structure generation and multitask GNN property prediction to accelerate discovery and overcome traditional tradeoffs.

The work builds on decades of polymer dielectrics research where BOPP has dominated due to low loss and large Eg but fails at high temperatures. High-temperature polymers like PI (Kapton), PEEK, PEI, and FPE provide thermal stability but have lower Eg and U. Prior AI and informatics approaches have guided polymer design up to ~100 °C energy density, including co-design strategies, genetic algorithms, variational autoencoders, and databases such as PI1M. Reaction-aware virtual libraries (OMG, SMiPoly) generate synthetically accessible polymers but may not fully enforce constraints favoring high molecular weight formation. ROMP has produced some of the best high-T dielectrics reported. There is also literature on composites with semiconducting nanofillers improving high-T U, and on green solvent selection and environmentally benign synthesis pathways. The present work advances by tightly integrating synthesizability-aware virtual forward synthesis with multitask GNN property prediction tailored to Tg, Eg, and ε, and by extending screening to solubility in green solvents.

AI-driven design via polyVERSE: The approach comprises three steps: (1) chemical structure generation using reaction templates applied to commercially available small-molecule monomers; (2) property prediction via multitask graph neural networks (polyGNN) for Tg, Eg, and ε; and (3) screening using property thresholds and ranking. - Monomer pool: SMILES from ZINC and ChEMBL databases were curated for ready purchasability and standard reactivity; counterions/chirality removed; synthetic accessibility (SAscore) used to filter complexity, resulting in ~8 million unique monomers. - Reaction-aware generation: Handcrafted reaction templates encode transformations and monomer filters for each polymerization. Focus initially on ROMP (ring-opening metathesis polymerization), with one transformation (double-bond cleavage at reactive site) and filters enforcing: exactly one cyclic olefin group; ring sizes of 3–5 or 7–11 atoms (to ensure strain and high conversion); exclude strong electrophiles adjacent to the olefin (e.g., halides, acyl halides, carbonyls, carboxyls) that reduce/distribute ring strain. The generation assumes perfectly repeating units and high molecular weight (a pragmatic approximation for narrow-pathway reactions). Result: 26,858 ROMP-based polymer structures. - Property prediction with polyGNN: Polymer repeat units converted to graphs (atoms as nodes, bonds as edges), and descriptors learned via neural message passing. Three multitask models trained: Tg (with melting temperature, decomposition temperature, thermal conductivity as auxiliary tasks); Eg (DFT-trained at crystal level with electron affinity and ionization energy as auxiliary tasks); ε at 100 Hz and room temperature (twelve-task model including ε at multiple frequencies, DFT zero-frequency ε, and refractive indices). Test-set RMSE: 32 °C (Tg), 0.5 (ε), 0.5 eV (Eg). - Screening: Apply thresholds Tg > 100 °C, ε > 3, Eg > 4 eV; rank remaining structures by product Tg × ε × Eg; select five top candidates based on cost and synthesizability for experimental validation. Synthesis and characterization: - Four polymers successfully synthesized via ROMP and cast into films: PONB-2Me5Cl, PNB-2,5DM, PNB-2Me5Cl, PNB-3Cl4Me; one candidate failed to polymerize. ROMP conditions included Grubbs-type catalysis, quench with ethyl vinyl ether, precipitation, purification, and drying. - Material characterization: NMR (Bruker 500 MHz) for structure and purity; TGA (TA Q-500) at 20 °C min−1; DSC (TA Q20) at 10 °C min−1 heating/cooling; GPC (Waters; DMAc mobile phase; PS standards) for molecular weight; UV-Vis (PerkinElmer Lambda 1050) for Eg. - Electrical characterization: Dielectric spectra via Solartron SI 1260/1296; temperature-controlled oven; sputtered Au/Pd electrodes (15 mm for spectroscopy; 3 mm for D–E loops). High-field D–E loops using Sawyer–Tower setup at 100 Hz, Trek 10/40 amplifier. Green-solvent-enabled screening and polyimide generation: - Trained a 61-task polyGNN classifier to predict polymer solubility (soluble/partial/insoluble) at room temperature across 61 solvents (including water, ethanol); softmax outputs; trained with class-weighted cross-entropy (Adam; Xavier init). Validation: overall F1 0.724, accuracy 89.7% on held-out; for unseen-polymers subset (800 pairs): F1 0.738, accuracy 94.9%; ethanol (30 pairs): F1 1.0, accuracy 100%; water (29 pairs): F1 0.759, accuracy 75.4%. - Extended polyVERSE with polyimide template (dianhydride + diamine to polyamic acid, then imidization). Monomer filters require exactly two functional groups with similar reactivity (proxied by Gasteiger charges). SMARTS transformations implemented; reject candidates with <8-atom shortest paths across repeat-unit edges to minimize backbiting. Generated 66,103 polyimide structures; screened for Tg > 200 °C, ε ≥ 3, Eg > 4 eV, and predicted solubility in water or ethanol; selected four candidates with affordable monomers and provided synthetic routes. Data/code availability: Generated structures and code for polyVERSE and polyGNN models are publicly available at the provided GitHub repositories.

- Discovery: A previously unknown polynorbornene dielectric, PONB-2Me5Cl, was discovered in silico and experimentally realized. At 200 °C, it delivers discharged energy density Ue = 8.3 J cm−3, approximately 11× any commercial polymer at this temperature, extending high-performance operation across 85–200 °C. - Breakdown and band gap: PONB-2Me5Cl exhibits Eg = 4.4 eV and extraordinary breakdown fields: >800 MV m−1 at 100 °C and ~750 MV m−1 at 200 °C. Room-temperature breakdown surpasses BOPP despite BOPP’s larger Eg, suggesting breakdown mechanisms in BOPP include non-electronic effects. - Dielectric behavior: Moderate, thermally stable ε with reasonable loss across frequency and temperature, sustaining high U from RT to 200 °C. - Comparative performance: Among discovered ROMP polymers, subtle substituent and backbone changes significantly affect Ebd and U. For instance, changing only an aryl substituent (PNB-2Me5Cl vs. PNB-2,5DM) alters Ebd by ~−30 MV m−1 and U by ~0.7 J cm−3 at 200 °C; modifying backbone (PNB-2Me5Cl vs. PONB-2Me5Cl) increases Ebd by ~300 MV m−1 and U by >5.5 J cm−3 at 200 °C. - ML prediction accuracy: For four synthesized polymers, mean absolute errors were 12 °C (5% MAPE) for Tg, 0.4 eV (8%) for Eg, and 0.7 (22%) for ε. - Library scale and screening: 26,858 ROMP structures generated; top candidates selected with Tg > 100 °C, ε > 3, Eg > 4 eV and ranked by Tg×ε×Eg; four successfully synthesized. - Environmental and synthesis context: PONB-2Me5Cl synthesis requires three steps (two to monomer, one polymerization), comparable to o-POFNB and fewer than PSBNP-co-PTNI0.02 (six steps). Eliminating cooling needs via high U can reduce system mass/volume. - Green solvent screening: Initial ROMP set yielded no candidates soluble in water/ethanol while meeting high-T dielectric criteria. Polyimide expansion (66,103 structures) identified several hundred predicted hits satisfying Tg > 200 °C, ε ≥ 3, Eg > 4 eV, and solubility in water or ethanol; four affordable candidates proposed. Predicted polyimides show slightly lower Tg (204–213 °C) than ROMP polymers (220–243 °C) but higher Eg (5.4–5.7 eV), implying potentially higher Ebd.

The study addresses the longstanding tradeoff between high-temperature stability and energy density in polymer dielectrics by integrating synthesizability-aware structure generation with accurate ML property predictions. PONB-2Me5Cl demonstrates that a carefully engineered backbone (bicyclic rings, double bonds for stiffness) and pendant groups (π-systems and polar groups) can simultaneously yield high Tg, moderate ε with low loss, and high Eg leading to exceptional Ebd, thereby maximizing U at elevated temperatures. The superior high-T performance relative to BOPP and other commercial polymers suggests that thermal and electromechanical breakdown pathways can be mitigated through targeted chemistry. Substituent-level and backbone-level modifications substantially impact Ebd and U, highlighting a sensitive structure-property landscape that polyVERSE can navigate to optimize performance. Environmentally, the high U and thermal stability could reduce or eliminate cooling systems, enabling lightweighting and lower energy consumption; synthesis-step analysis suggests PONB-2Me5Cl’s impact is competitive with comparable high-T ROMP dielectrics. Recognizing limitations of solvent compatibility, the solubility model and polyimide expansion offered a path to high-T dielectrics compatible with water/ethanol, balancing high Eg (and thus Ebd potential) with adequate Tg. Collectively, these findings validate AI-driven workflows as practical tools for rapid discovery and optimization of high-T dielectrics and suggest promising extensions (e.g., nanofillers, R-group engineering, greener processing) to further elevate performance and sustainability.

This work introduces the polyVERSE paradigm for AI-assisted polymer discovery, combining reaction-aware virtual synthesis from commercially available monomers with multitask GNN property prediction and targeted screening. The approach yielded PONB-2Me5Cl, a high-temperature dielectric achieving Ue = 8.3 J cm−3 at 200 °C (≈11× commercial options), with exceptional breakdown strength and stable dielectric properties across temperature. The methodology uncovered the strong influence of subtle chemical modifications on Ebd and U, enabling rational tuning within polynorbornenes. Extending the pipeline with a solubility model and a polyimide reaction template identified water/ethanol-compatible candidates with high Eg and adequate Tg, pointing toward greener processing. Future work should: explore additional polymerization templates to broaden chemical space; systematically study structure–breakdown causality; integrate nanofillers/coatings and R-group engineering to further raise U and reduce loss; and experimentally validate the predicted green-solvent-soluble polyimides. The demonstrated data/code availability facilitates community adoption and accelerates further advances.

- Modeling approximations: Property models assume perfectly repeating units and effectively infinite molecular weight; ε predicted at room temperature and 100 Hz only, whereas ε is temperature- and frequency-dependent. Breakdown is not directly simulated under engineering conditions; Eg is used as a proxy for Ebd. - Chemical space coverage: Initial generation focused on ROMP; omission of certain monomers (e.g., for PONB-2,5DM) due to database versioning demonstrates dependency on monomer catalogs. One of five selected candidates failed to polymerize. - Understanding of mechanisms: Precise reasons for large Ebd changes from subtle substituent/backbone variations are not fully resolved; BOPP’s lower Ebd despite larger Eg suggests additional, not fully quantified, thermal/electromechanical effects. - Solubility model performance: While overall high accuracy, performance for water with unseen polymers is lower (F1 ≈ 0.759; accuracy ≈ 75.4%), introducing uncertainty in green-solvent screening. - Experimental scope: High-field performance characterized for select candidates; broader validation and long-term reliability (e.g., cycling, humidity, mechanical durability) are not detailed. Green processing for ROMP polymers remains limited as PONB-2Me5Cl is insoluble in water/ethanol and common green solvents.