Brett Matthew Savoie

Memory systems for AI agents are a valuable component of long-running workflows. They are also a strong form of lock-in for model providers, because the memory system that makes a model better with use may be inaccesible to users. An ideal memory system would accumulate both factual knowledge and the relationships of that knowledge to recurring tasks performed by the researcher. People accumulate particular and general knowledge through experience, the latter often through periodic reflection on what is working and isn’t working at a higher level of abstraction. Here, we introduce an open-source harness for polymer research meant to heuristically mimic (i) knowledge accumulation across research tasks, (ii) the review of new knowledge against existing knowledge, and (iii) performing periodic updates to task-knowledge relationships through reflection. The system, which we call the Open Ontology of Polymer Chemistry (OOPC, pronounced ‘oopsie’), is initialized with a knowledge base consisting of an open-source wikipedia with entries for most known polymers. From this starting point, the knowledge base was enriched by running a 1,092 task campaign across different areas of polymer chemistry, producing 8,899 distinct findings extracted from 1,679 sources on 157 different polymers. GPT-6.1 Sol and Opus 5.5 comparisons were run with and without the accumulated memory from this campaign on 60 held-out tasks with varying overlap in terms of polymer and properties being researched. The performance improvements tracked with research topic similarity to prior work researched from the harness, with token reductions of up to 19% and 48% and web search reductions of up to 80% and 72% (10.3 to 2.1 and 7.6 to 2.1 per answer) for Opus 5.5 and GPT-6.1 Sol, respectively. The harness was tested through five independent restarts of 500 tasks each with reproducible performance gains. An independent review of the performance caught rare contradictory claims and occasionally unsubstantiated claims depending on the underlying model (up to a few percent of all claims by claude on withheld tasks, none by codex). The reflection heuristic was tested as an update to the knowledge ontology after the accumulation period. These updates mainly consisted of extended lists of property names with marginal improvements to finding recall. The roles of review and ontology updates require further study as after knowledge accumulation these tests had little room to show appreciable gains given the consistently strong evidentiary basis for findings within the harness. Looking forward, domain-specific memory systems are likely to be valuable and ideally transferable research artifacts that can be shared at the group, community, and organizational levels.

High‐concentration electrolytes (HCEs) are investigated for their ability to form robust, inorganic‐rich solid electrolyte interphases (SEI). While the SEI composition in HCEs differs from that in dilute electrolytes, the specific anion‐mediated reaction network driving this transformation remain poorly understood. The network complexity and transient nature of the radical intermediates involved have precluded the development of a comprehensive mechanistic map of initial SEI formation. Here, the reaction networks governing anion‐mediated electrolyte degradation are established using automated reaction discovery and microkinetic modeling (MKM). The study encompasses the decomposition pathways of ethylene carbonate (EC), propylene carbonate (PC), dimethyl carbonate (DMC), and acetonitrile (AN) in the presence of reduced anion species derived from LiFSI and LiTFSI. The results reveal a kinetically accessible landscape for solvent decomposition driven by fluoride () and fragments, identifying specific pathways that favor the formation of inorganic LiF and open‐ring organic carbonates. Microkinetic modeling further indicates that acts as a rapidly consumed driver, shifting the product distribution toward thermodynamically stable fluoro‐organic species. These results provide a molecular‐level rationalization for the unique SEI chemistries observed in HCEs and suggest that tuning the anion‐solvent interaction is a viable pathway for controlling interphase growth.

Low-temperature oxidation reactions are central to fuel and combustion chemistry, yet the transient radical intermediates that govern them remain difficult to interrogate experimentally. The oxidation of tetrahydrofuran (THF) to ketohydroperoxide (KHP) is a paradigmatic system in which established mechanistic proposals fail to reproduce experimentally observed product distributions. Six KHPs are accessible through THF oxidation, three following molecular oxygen addition to the α carbon and three following addition to the β carbon. Although α-addition is both kinetically and thermodynamically preferred and has been characterized in detail, the experimental β-derived KHP abundances remained incompletely explained by existing models. Here the THF-to-KHP network is mapped with an automated reaction exploration algorithm that computes activation energies and reaction thermodynamics for all KHP-forming pathways and for potentially competitive off-target channels. The exploration recovers previously reported reactions in quantitative agreement with the literature while identifying hydrogen-transfer and isomerization reactions that connect the α and β branches of the network. The energetic accessibility of these interconnecting reactions is consistent with the reported abundance of β-pathway KHPs and identifies specific channels that should be included in future kinetic models of THF oxidation.

Mechanistic modeling of chemical transformations offers a compelling basis for understanding reactivity and allows for prediction of reaction outcomes before attempting experiments. Despite progress in machine-learned interatomic potentials (MLIPs), we demonstrate that available models lack the accuracy for diverse reaction modeling. With this motivation, we developed a general MLIP for mechanistic modeling of closed-shell carbon, hydrogen, nitrogen, and oxygen reactions, AIMNet2-rxn, using a dataset of ∼4.7 × 10 ⁶ range-separated density functional theory calculations. AIMNet2-rxn enables reaction modeling ∼10 ⁶ faster than the reference quantum mechanical (QM) methods while substantially outperforming graph-based ML, reaffirming the value using three-dimensional chemical information for training. On a test suite of well-known reaction mechanisms—such as amide formation, proton transfers, and pericyclics—AIMNet2-rxn yields 1 to 2 kilocalories per mole accuracy across reaction coordinates without retraining or system-specific fine-tuning. To exploit graphics processing unit parallelism and AIMNet2-rxn efficiency, we introduce a batched nudged elastic band procedure that readily achieves minimum energy pathway search on a millions-of-reactions scale. To demonstrate complex reaction characterization, the thermodynamics of an 11-step pathway producing hydroxymethylfurfural, the experimentally observed major product of glucose pyrolysis, is evaluated. Overall, the accuracy and efficiency afforded by AIMNet2-rxn create opportunities in high-throughput reaction discovery and deep reaction network analysis that would be infeasible with QM methods.

High-concentration electrolytes (HCEs) improve lithium-ion battery cycling by shifting interphase formation from the solvent to the anion, producing a more stable inorganic-rich interphase at the anode. However, the anions for HCEs has been mainly studied on FSI− and TFSI−, and the anion itself has received little systematic exploration relative to the solvent and the salt concentration. Here, anion chemical space is searched using evolutionary framework that couples genetic algorithm with quantum-chemical descriptors of reducibility, solvent coordination, and charge delocalization. Substructure analysis of the evolved populations identifies a sulfinimide (N−=S) core as the recurring motif in place of the sulfonyl chemistry of the reference anions, and a combinatorial scaffold-branch library built from these motifs exhibits distinction between the electronic and intermolecular contributions to anion performance. Charge localization is governed by the scaffold while solvent binding is set neither by scaffold nor by branch alone but by their specific pairing and by whether the two arms are chemically distinct. Higher-level solvation show that the leading candidates from the selected anion set exceed both baseline anions in bulk solvation, and transport calculations exhibit that every candidate carries a smaller Li+ first shell and diffuses faster than both baselines at concentrated salt content. Replacing a single sulfonyl oxygen with an imide nitrogen reorganizes the anion around a delocalized, asymmetrically substituted core and opens a salt chemistry for HCEs distinct from the sulfonyl imides in use today.

Chemical imputation models provide a flexible strategy for estimating missing values in sparse chemical databases.

Zip length, or the number of monomers released between initiation of depolymerization and loss of the depropagating radical species via termination and/or transfer, is a key metric for assessing the kinetic recyclability of a polymer, with higher values indicating greater potential for monomer recovery. In the context of radical depolymerization, mechanism-dependent analytical zip length expressions have been derived as a function of kinetic rate coefficients, molecular weight, density, and temperature; continuum models based on the method of moments (MoM) reproduce these expressions because they share the same underlying assumptions. However, we demonstrate that these analytical expressions do not accurately reflect the true zip length under typical conditions due to the accumulation of monomer radicals that cannot undergo further depropagation. To investigate this phenomenon, we compared two kinetic modeling approaches, kinetic Monte Carlo (KMC) and population balance modeling with the MoM, across a broad parameter space spanning two initiation mechanisms, multiple kinetic parameter sets, and a range of initial chain lengths and dispersities. The relationship between initial chain length and simulated zip length is not a simple threshold comparison but rather depends on the ratio of the initial degree of polymerization to zip length, the initiation mechanism, and the kinetic parameter set, which we have quantified using a dimensionless value we term “scaled kinetics of zip length”, or the Skoz number. We then apply an expanded KMC model with kinetic parameters derived from DFT calculations to poly(diethyl muconate), a recently developed recyclable polyolefin system with a mechanism akin to those analyzed, demonstrating how a mechanistically grounded understanding of zip length, including its evolution during depolymerization and its dependence on initiation barrier, can guide the rational design of kinetically recyclable polymers.

Organic mixed ionic–electronic conductors (OMIECs) are π-conjugated polymer–electrolyte composites that support coupled ionic and electronic transport, making them candidates for bioelectronics, electrochemical sensors, and energy storage devices. Despite growing interest in these materials, the joint influence of electrolyte composition and polymer side-chain design on OMIEC morphology and charge transport remains poorly understood, owing to the small number of material systems that have been experimentally or computationally characterized. Here, coarse-grained molecular dynamics (CG-MD) and kinetic Monte Carlo (kMC) simulations are used to systematically vary anion size (1–3 Å) and polymer side-chain design (polarity and grafting frequency) across nine OMIEC configurations. The simulations reveal that side-chain frequency is the dominant design lever for polaron mobility, with half-frequency intermediate polar side chains achieving approximately 40% higher mobility than full-frequency configurations owing to a significant reduction in trap-site formation. Smaller anions percolate more readily near polymer side chains, with 1 Å anions achieving roughly 25–50% higher local concentration than 3 Å anions across the studied systems, but this enhanced percolation does not translate into significant differences in backbone organization or mobility. Reduced side-chain frequency also promotes longer-range π–π stacking extending to approximately 14 Å, while anion size and side-chain hydrophilicity exert minimal influence on backbone alignment. These results identify side-chain frequency as a more effective optimization target than anion size or hydrophilicity for improving OMIEC conductivity and provide a computational framework for establishing quantitative design rules across a broader range of mixed conducting polymers and electrolytes.

Radical polymers are an emerging class of organic electronic materials that offer electronic properties, making them attractive for transparent, biosensing, and energy storage applications. Charge transport in these materials occurs through redox-active side chains rather than a conjugated backbone. However, their practical use is limited by relatively low charge mobility compared to conjugated polymers, in which charge is transported along a continuous conjugated backbone. In addition, most designs have focused on TEMPO-based systems and small molecules, leaving the broader design space of radical chemistries and backbone structures insufficiently explored. Here, a multiscale modeling framework is used to systematically investigate the influence of radical side-chain chemistry and polymer backbone structure on charge transport in radical polymers. The approach integrates system-specific coarse-grained parametrization derived from machine-learned interatomic potentials with molecular dynamics simulations of equilibrated morphologies and kinetic Monte Carlo charge transport calculations. The effects of radical side-chain chemistry, polymer backbone structure, and active radical-site fraction were systematically investigated across polymers containing TEMPO, verdazyl, and galvinoxyl side chains with PEO and PMMA backbones. The results show that radical side-chain chemistry strongly influences both morphology and charge transport, with galvinoxyl exhibiting the highest charge mobility, followed by verdazyl, while TEMPO systems exhibit comparatively lower mobility. Galvinoxyl achieves high mobility through large intersite hopping rates and the formation of extensive connected charge-transport networks, whereas verdazyl balances intermediate hopping rates with high site connectivity. In contrast, TEMPO-based systems exhibit limited transport because unfavorable intermolecular hopping kinetics dominate despite their higher site connectivity relative to galvinoxyl. The coarse-grained representation of PEO and PMMA backbones produces similar averaged mobilities, suggesting that backbone-dependent transport differences observed experimentally may originate from atomistic effects not captured by the CG model. Computational annealing does not produce significant changes in radical packing or charge transport, likely because equilibrated CG morphologies are already partially relaxed due to the initial equilibration procedure. Finally, reducing the fraction of active radical sites to mimic synthesis imperfections, such as protonation and missing side chains, decreases charge displacement and mobility, while the resulting changes in transport-network formation depend on the underlying hopping kinetics and network topology. These findings establish structure--property relationships governing charge transport in radical polymers and provide design guidelines for improving their conductivity. More broadly, the workflow establishes a general framework for probing mesoscale structure--transport relationships using system-specific force fields and is readily extensible to a broader molecular design space.