Source-linked AI summary
ElectrolyteFM: Unifying Electrolyte Property Prediction through Cross-Property Knowledge Learning
Jiaxin Yu, Shuo Wang, Peng Wang, Yongcai Wang, Deying Li
TL;DR
Electrolyte property predictors often cover a limited set of properties, although formulation design requires balancing multiple observables and cross-property supervision can help or interfere. ElectrolyteFM combines independently learned property-specific representations with selectively routed shared expertise and residual fusion. Across 12 properties, it reduces macro NMAE by 14.8% relative to the strongest electrolyte-specific baseline.
Problem
Existing electrolyte predictors often focus on a limited set of properties, while heterogeneous properties may benefit differently from cross-property supervision.
Source: §1 · p. 1Method
ElectrolyteFM independently learns property-specific representations and shared expertise, then routes relevant shared information for each formulation and target property and fuses it through residual adapters.
Source: §1 · pp. 2–3Results
Across 12 electrolyte properties, ElectrolyteFM reduces macro NMAE by 14.8% relative to the strongest electrolyte-specific baseline.
Source: §1 · p. 3Takeaways & Limitations
ElectrolyteFM achieves the lowest macro NMAE and lowest MAE on 10 of 12 properties, with conductivity OOD and few-shot results supporting prediction and adaptation under formulation shift.
Source: §5 · p. 11
Abstract
from arXiv · showhide
Electrolyte formulation design requires balancing multiple physicochemical properties, yet existing models often focus on a limited subset. Learning each property in isolation can overlook transferable chemical information, whereas indiscriminate sharing can introduce cross-property interference. Our directed transfer analysis shows that jointly learning two property prediction tasks can improve or degrade prediction relative to separate training, with asymmetric transfer effects between the tasks. We propose ElectrolyteFM, a unified multi-property prediction model which can more accurately predict multiple properties of each electrolyte by effectively identifying and utilizing property-specific features and knowledge shared across properties. More specifically, ElectrolyteFM learns property-specific representations independently and captures cross-property knowledge through a separately trained expert pool. A router selects relevant shared information for each formulation and target property, and property-specific residual adapters convert this information into corrections to the corresponding representation for prediction. Experiments on Electrolyte12 show that ElectrolyteFM reduces normalized mean absolute error averaged across 12 electrolyte properties by 14.8% relative to the strongest electrolyte-specific baseline. On an independent sodium-electrolyte dataset unseen during training, it reduces conductivity mean absolute error by 6.7% relative to the best-performing baseline.
1. Introduction
Electrolyte formulation design requires balancing multiple properties, but existing predictors often cover only a limited subset and cross-property supervision can help or interfere. ElectrolyteFM combines property-specific representations with selectively routed shared knowledge to predict across properties.
1. Introduction
- Electrolyte formulation design is a multi-objective problem, while existing predictors often cover only a limited set of properties despite their shared formulation and interaction basis.A unified model could reuse information across datasets and provide a more comprehensive representation of electrolyte behavior.Source: §1 · p. 1
- Across 12 properties, directed transfer includes beneficial, harmful, and asymmetric interactions, motivating selective sharing that preserves property-specific knowledge.The analysis reports 61 positive and 71 negative transfer scores among 132 directed scores.Source: §1 · pp. 2–3
- ElectrolyteFM learns property-specific representations independently and uses a separately trained expert pool, formulation–property-conditioned routing, and property-specific residual adapters to fuse shared information.The adapters convert routed shared information into corrections to the corresponding private representations.Source: §1 · pp. 2–3
- The study curates Electrolyte12 and evaluates multi-property prediction, learning strategies, knowledge-branch contributions, and zero-shot generalization and few-shot adaptation on an independent sodium-electrolyte dataset.The sodium evaluation reports conductivity MAE improvement relative to the best-performing baseline evaluated on that dataset.Source: §1 · p. 3
- 14.8% lower macro NMAE than the strongest electrolyte-specific baseline; on an independent sodium-electrolyte dataset, conductivity MAE is 6.7% lower than the best-performing baseline evaluated there.The sodium-dataset result also reports further gains from few-shot adaptation.Source: §1 · p. 3
2. Related Work
Prior work develops molecular and formulation representations, transfer-learning strategies, and mixture-of-experts approaches. ElectrolyteFM addresses how shared knowledge contributes to each formulation–property prediction by combining separately learned private and shared representations.
2.1. Learning for Electrolyte Formulations
- Molecular and formulation-level models improve electrolyte representations, but these advances alone do not resolve how heterogeneous property supervision should be shared.ElectrolyteFM combines property-specific learning with selective sharing for unified multi-property prediction.Source: §2.1 · p. 3
2.2. Cross-Task Transfer Learning
- Task selection, grouping, and gradient coordination address which tasks share supervision and how joint optimization is regulated, while ElectrolyteFM routes shared information for each formulation and target property.This routing is motivated by directed transfer analysis.Source: §2.2 · p. 4
2.3. Mixture-of-Experts for Knowledge Sharing
- ElectrolyteFM separately learns property-specific private representations and a shared expert pool, then freezes both branches for fusion through property-specific residual adapters.The adapters turn routed shared information into corrections to private representations.Source: §2.3 · p. 4
3. Method
The method measures directional transfer between electrolyte properties, then combines independently learned property-specific representations with selectively routed shared knowledge. Its three learning stages preserve private patterns while adapting shared information to each formulation and target property.
3.1. Directed Cross-Property Transfer Analysis
- Directed transfer is heterogeneous and asymmetric: supervision from one property can help some targets while interfering with others, motivating selective sharing.The analysis compares target-only and pairwise jointly trained models on the target property; a positive transfer score denotes improvement and a negative score interference.Source: §3.1 · pp. 4–5
3.2. Problem Formulation and Overview
- ElectrolyteFM combines property-specific and cross-property knowledge through private experts, formulation–property conditioned routing, and residual adapters that correct private representations.The problem setting has 12 properties and partially observed labels; objectives use transformed and standardized targets with z-MSE.Source: §3.2 · p. 5
3.3. Specific Knowledge Learning
- Specific Knowledge Learning assigns each property an independent expert and prediction head, trained only on that property’s data with the encoder frozen.Each private expert produces a property-specific state and standardized prediction.Source: §3.3 · p. 6
3.4. Shared Knowledge Learning
- Shared Knowledge Learning uses a pool of shared experts and a formulation–property conditioned router to select and weight reusable information for each prediction.ShKL trains the shared encoder, experts, router, and prediction heads jointly while keeping the molecular graph module fixed; the learned encoder, experts, and router are retained for fusion.Source: §3.4 · p. 6
3.5. Shared-to-Specific Knowledge Fusion
- Shared-to-Specific Knowledge Fusion applies a property-specific residual adapter to turn routed shared information into a correction to the private representation.The fused model starts from the learned specific predictor; during fusion, the specific and shared branches are frozen while the adapter and prediction head are optimized.Source: Figure 2 · p. 5, §3.5 · pp. 6–7
4. Experiments
ElectrolyteFM performs strongly across Electrolyte12 and conductivity prediction on unseen sodium formulations. The experiments also test zero-shot prediction and few-shot adaptation under formulation shift.
4.1. Experimental Setup
- The evaluation uses Electrolyte12, with shared train, validation, and test splits across methods and both experimental and computed labels.The test set includes 43,709 records; the training and validation sets contain 366,825 and 44,811 records, respectively.Source: §4.1 · p. 7
4.2. Multi-Property Prediction
- ElectrolyteFM achieves the best MAE on 10 of 12 properties and the lowest macro NMAE, 0.1571 versus 0.1845 for the strongest baseline.It wins on all four experimentally labeled properties and six of eight computed-label properties.Source: §4.2 · p. 7, Table 1 · p. 8
4.3. Generalization to Unseen Formulations
- Generalization tests use 106 independent experimental sodium-electrolyte formulations, comparing direct zero-shot prediction with adaptation using a small number of target-domain labels.The formulations include mixtures with up to 14 components.Source: §4.3 · p. 8
4.3.1. Zero-shot Generalization
- ElectrolyteFM ranks first across all three metrics in both ID and OOD settings, reducing OOD conductivity MAE by 6.7% relative to SCAN.The baseline ranking changes under distribution shift: Bamboo-Mixer beats SCAN on ID data but falls behind it on OOD data.Source: §4.3.1 · p. 8, Table 2 · p. 8
4.3.2. Few-shot Adaptation
- ElectrolyteFM has the lowest mean query z-MSE at every evaluated support size under both adaptation protocols.With 16 labeled formulations and fixed learned representations, its mean query z-MSE falls from 0.2552 to 0.2423 with affine calibration and 0.2373 with final-layer adaptation; Figure 3 plots the comparison.Source: §4.3.2 · p. 9, Figure 3 · p. 9
4.4. Ablation Studies
ElectrolyteFM combines independent property-specific learning with selectively routed shared knowledge. Its aggregate results and ablations favor this combination over the compared alternatives and show that expert allocation and predictive contribution are not identical.
4.4.1. Comparison of Learning Strategies
- ElectrolyteFM reduces macro NMAE by 2.5% relative to PLE and has the highest macro PCC and R^2 among the compared learning strategies.The multi-task baselines beat STL on all three macro metrics, but their improvements over STL occur in the experimental-label group and reverse in the computed-label group; ElectrolyteFM improves all three metrics in both groups.Source: §4.4.1 · p. 9, Table 3 · p. 10
4.4.2. Contribution of Knowledge Learning and Fusion
- The shared branch alone outperforms the private branch, while the complete model lowers macro NMAE by 14.7% versus the private branch and 6.1% versus the shared branch.The results support complementary information from the private representation and routed shared representation combined through property-specific residual correction.Source: §4.4.2 · p. 10, Table 4 · p. 10
4.4.3. Expert Allocation and Predictive Contributions
- Mean routing preferences vary by property, and removing any expert increases macro z-MSE; removing S1 has the largest effect at 12.43%.S2 contributes despite having no property's largest mean routing weight; Figure 4 shows mean weights and the relative error increase after expert removal.Source: §4.4.3 · pp. 10–11, Figure 4 · p. 10
5. Conclusion
ElectrolyteFM separates private and shared representations, then uses property-specific residual adapters to combine them. The paper concludes that the model performs well across electrolyte properties and supports prediction under formulation shift.
5. Conclusion
- ElectrolyteFM learns private and shared representations separately and integrates them through property-specific residual adapters.The paper presents this design as a response to heterogeneous cross-property transfer.Source: §5 · p. 11
- ElectrolyteFM achieves the lowest macro NMAE and the lowest MAE on 10 of the 12 properties.Source: §5 · p. 11
- Conductivity OOD and few-shot results further support the model's use for prediction and adaptation under formulation shift.Source: §5 · p. 11
A. Additional Multi-Property Prediction Results
Additional evaluation reports correlation and R^2 across Electrolyte12 properties. ElectrolyteFM ranks first for both metrics on nine properties and has high macro scores.
A. Additional Multi-Property Prediction Results
- ElectrolyteFM ranks first on nine properties for each metric, covering all four experimental-label targets and five computed-label targets.Tables 5 and 6 report per-property PCC and R^2, respectively.Source: A. Additional Multi-Property… · p. 15, Table 5 · p. 15, Table 6 · p. 15
- ElectrolyteFM's macro PCC is 0.9268.Source: A. Additional Multi-Property… · p. 15, Table 5 · p. 15
- ElectrolyteFM's macro R^2 is 0.8357.Source: A. Additional Multi-Property… · p. 15, Table 6 · p. 15
B.1. Comparison of Learning Strategies
ElectrolyteFM ranks best on macro error and association metrics across both experimental-label and computed-label groups.
B.1. Comparison of Learning Strategies
- ElectrolyteFM has the lowest macro NMAE in each label-source group.Source: §B.1 · p. 16
- ElectrolyteFM has the highest macro PCC and R2 in each label-source group.Source: §B.1 · p. 16
- The comparison averages metrics equally over four experimental-label and eight computed-label properties.Bold and underline mark the best and second-best values within each column.Source: Table 7 · p. 16
B.2. Contributions of Knowledge Learning and Fusion
Across twelve properties, S2SKF outperforms SpKL and ShKL on macro NMAE and leads on macro PCC and R2.
B.2. Contributions of Knowledge Learning and Fusion
- The ablation reports macro NMAE, PCC, and R2 for configurations in Table 4; SpKL and ShKL use their own prediction heads, while S2SKF combines both learned representations.Check marks indicate included components, and the metrics give equal weight to each property.Source: §B.2 · p. 16, Table 8 · p. 16
- S2SKF reduces twelve-property macro NMAE by 14.7% versus SpKL and 6.1% versus ShKL.Source: §B.2 · p. 16
- S2SKF achieves the highest macro PCC and R2 across all twelve properties.The reported comparison describes aggregate performance as improved across all three metrics.Source: §B.2 · p. 16
B.3. Expert Allocation and Removal
Expert-removal tests show that every expert contributes to twelve-property macro z-MSE, with S1 removal having the largest and label-group-dependent effect.
B.3. Expert Allocation and Removal
- The removal test zeros one expert’s routing weight, renormalizes the remaining weights, and evaluates the resulting shared state using the same frozen adapter and head.The evaluation uses the training target transforms and standardization statistics.Source: §B.3 · p. 16
- Removing any expert increased twelve-property macro z-MSE; removing S1 produced the largest increase (+12.43%), with +67.35% for experimental-label versus +4.65% for computed-label properties.Source: §B.3 · p. 16, Table 9 · p. 17
- Average routing preference and removal cost provide complementary views of expert use, whose contributions vary across evaluated properties.Source: §B.3 · p. 17