Comprehensive Workflow for Uncertainty-Aware Ionic Liquid Solvent Discovery

Alexander Lobo; Urvi Awasthi; Leonid Zhukov

cond-mat.mtrl-sciOct 12, 2026

Comprehensive Workflow for Uncertainty-Aware Ionic Liquid Solvent Discovery

Authors: Alexander Lobo, Urvi Awasthi, Leonid Zhukov

Organizations: BCG X AI Science Institute

Abstract

Designing synthesizable materials with targeted properties remains one of the central objectives in materials sciences. Ionic liquids (ILs) are well suited to this goal due to their "tunability", where physical and chemical properties can be precisely customized by modifying their molecular structure, specifically by varying the combination of cations and anions. At the same time, current data-driven property models for ionic liquids often degrade in scarce, noisy, and uneven data regimes, where measurements cover only a small fraction of chemical space and a limited range of thermodynamic conditions. This makes out-of-distribution screening brittle: even when point predictions look plausible, it is difficult to compare candidates consistently, quantify downstream implications, or trust rankings enough to guide molecular discovery. We present a property-prediction and optimal candidate selection workflow that treats data processing, deep learning modeling, and physics-based calculations as a coupled system, linking property prediction with uncertainty estimation and physics-based filtering. By tightening the link between predicted properties and the criteria used for down-selection, the workflow supports reliable out-of-distribution ranking and enables prioritization of previously untested ionic liquid candidates with favorable predicted trade-offs. We demonstrate the application of our workflow by designing novel ionic liquids for CO2 capture from flue gas, predicting their properties and selecting Pareto-optimal candidates based on CO2 solubility, IL viscosity and working capacity.

Figures & tables

Fig. 1: Uncertainty-aware workflow for ionic liquid solvent discovery. In (A) , the MLP is trained as a five-member resampled ensemble, each member defining one predictive “world.” In (C) , working capacity $\Delta C$ and the heat-of-absorption proxy $-\Delta H$ are derived from a Van’t Hoff fit of the predicted temperature-dependent solubility curve. In (D) , candidates are filtered by objective-level feasibility cutoffs, then ranked by Pareto non-dominance probability $P_{\mathrm{ndom}}$ across the five worlds.

	Viscosity $\ln(\eta)$	CO 2 solubility $x_{\mathrm{CO_{2}}}$
Sources	ILThermo [ Dong2007 ] + literature [ VENKATRAMAN2017162 , CHEN2025335 , SONG2020115752 ]
$N$ (measurements)	20,309	17,261
Unique ILs	2,744	414
Cations / anions	1,218 / 315	159 / 107
$T$ range (K)	$[253,420]$	$[243,420]$
$P$ range (bar)	— (ambient)	$[0,499.9]$

TABLE I: Dataset summary for the viscosity and CO 2 solubility prediction targets, after curation. Filtering criteria are applied during curation (Appendix VII ).

Fig. 2: Three-branch MLP architecture for IL property prediction. Descriptor, Morgan-fingerprint, and thermodynamic-condition inputs are each processed by a separate subnetwork, then concatenated and passed through a shared head to predict the targets

\hat{y}

\ln(\eta)

for viscosity or

\mathrm{logit}(x_{\mathrm{CO_{2}}})

for solubility.

Model	CV MAE	CV $R^{2}$	Test MAE	Test $R^{2}$
XGBoost	$0.48_{\pm 0.08}$	$0.74_{\pm 0.10}$	$0.51$	0.81
Random Forest	$0.61_{\pm 0.03}$	$0.86_{\pm 0.06}$	$0.63$	0.74
ExtraTrees	$0.65_{\pm 0.01}$	$0.89_{\pm 0.03}$	$0.71$	0.72
Chemprop	$0.61_{\pm 0.12}$	$0.87_{\pm 0.20}$	$0.53$	0.69
ChemBERTa (finetune)	$0.64_{\pm 0.07}$	$0.91_{\pm 0.09}$	$0.70$	0.71
ILTransR (finetune)	$0.46_{\pm 0.07}$	$0.84_{\pm 0.06}$	$0.51$	$0.78$

TABLE II: Baseline predictive performance. Baseline performance under 5-fold balanced pair-scaffold cross-validation (CV) and pair-scaffold-disjoint held-out test evaluation for both targets:

\ln(\eta)

and

x_{\mathrm{CO_{2}}}

. CV metrics are computed from pooled out-of-fold predictions across the five folds. CV metrics are reported as mean

\pm

standard deviation across the five folds; held-out test metrics are reported as single evaluations on the pair-scaffold-disjoint test set.

Fig. 3: Distributions of predicted properties. CO 2 solubility and natural log viscosity properties for generated ionic liquids compared to those from raw training dataset. Plots show property distributions at absorption conditions ( $T=313.15\text{K},P=1\text{bar}$ ). Pareto front distribution represent isolated top performers and a subset of generated.

Fig. 4: Probabilistic Pareto front. Working capacity versus ln-viscosity for filtered IL candidates ( $N\approx 26{,}000$ ). Color indicates regeneration energy (kJ/mol CO 2 ). Orange circles highlight candidates appearing on the probabilistic Pareto front. Circle intensity is proportional to the Pareto non-dominance probability, $P_{\mathrm{ndom}}$ , reflecting the fraction of ensemble worlds in which a candidate is Pareto-optimal. Labeled candidates are top-ranked ILs (see Table III )

Panel of Figure 7 — Fig. 5: Family-level aggregation of filtered ionic liquid candidates, grouped by cation and anion families. Stars indicate family combinations with at least one candidate appearing on the probabilistic Pareto front ( $P_{\mathrm{ndom}}\geq 0.4$ ). Bubble size is proportional to the number of candidate ILs within each family combination. To avoid bias toward families with larger candidate pools, values are computed using only the top 50 performers per family (for respective metric). Arrow indicates whether higher or lower is better. ( a ) Average working capacity, highlighting families that achieve large absorption–desorption swings. ( b ) Average ln-viscosity, highlighting families that favor fluid operability under absorption conditions.

Rank	IL ID/Name	$P_{\mathrm{ndom}}$	$\Delta C$ $\uparrow$	$x_{\mathrm{CO}_{2}}^{\mathrm{abs}}$ * $\uparrow$	ln( $\eta$ ) $\downarrow$	- $\Delta H_{\text{abs,CO}_{2}}$ $\downarrow$
1	AmAm1	1.0	0.197 $\pm$ 0.068	0.607 $\pm$ 0.050	3.59 $\pm$ 0.32	$5.23$ $\pm$ 2.03
2	AmAm2	0.8	0.198 $\pm$ 0.055	0.629 $\pm$ 0.042	3.75 $\pm$ 0.22	$5.01$ $\pm$ 1.64
3	AmAm3	0.8	0.188 $\pm$ 0.059	0.613 $\pm$ 0.048	3.45 $\pm$ 0.33	$4.78$ $\pm$ 1.49
4	UnAm1	0.8	0.168 $\pm$ 0.063	0.653 $\pm$ 0.049	3.76 $\pm$ 0.18	$3.85$ $\pm$ 1.61
5	AmAm4	0.8	0.198 $\pm$ 0.061	0.622 $\pm$ 0.045	3.66 $\pm$ 0.29	$5.08$ $\pm$ 1.80
6	UnAm2	0.8	0.180 $\pm$ 0.062	0.545 $\pm$ 0.064	2.96 $\pm$ 0.27	$5.22$ $\pm$ 1.70

TABLE III: Selected Pareto front candidates. Top section shows top 10 anonymized high-confidence candidates (

P_{\mathrm{ndom}}\geq 0.6

); bottom section reveals structures for ranks 91–100 (

P_{\mathrm{ndom}}=0.2

). All metrics reported as mean

\pm

standard deviation across 5 ensemble worlds.

Category	Features
Element counts	num_B, num_C, num_N, num_O, num_F, num_Si, num_P, num_S, num_Cl, num_Br, num_I, num_Se
Aggregates	num_halogen (F+Cl+Br+I)
Molecular counts	formal_charge , heavy_atom_count

TABLE IV: Additional per-ion count features appended to the RDKit descriptor vector (15 features).

Dataset	$M$	$N$	mean $s$	median $s$	p95( $s$ )	mean $w_{90}$	p95( $w_{90}$ )
Viscosity (subsample 90%)	5	408634	0.238	0.210	0.487	0.520	1.067
Viscosity (bootstrap)	5	408634	0.272	0.235	0.570	0.595	1.246
CO 2 (subsample 90%)	5	408634	0.023	0.012	0.072	0.049	0.157
CO 2 (bootstrap)	5	408634	0.030	0.017	0.093	0.064	0.202

TABLE V: Summary of ensemble uncertainty distributions over the absorption-condition screening set (

N=408{,}634

candidate ILs). For viscosity, statistics are in

\ln(\eta)

units; for CO 2 solubility, in

x_{\mathrm{CO_{2}}}

units.

w_{90}

denotes the empirical ensemble 90% interval width

Q_{0.95}-Q_{0.05}

Materials Property Prediction

Materials property prediction uses data and physical models to estimate how composition, structure, and processing determine a material’s behavior.

See all 174 papers

A green solvent screening tool for emerging materials via uncertainty aware, transformer enhanced transfer learning
Jun 11, 2026Ioannis Kouroudis, Simon Ternes, Zhaosu Gu +5Molecular Property Prediction
Chair of Simulation of Nanosystems for Energy Conversion, Department of Electrical Engineering, TUM School of Computation, Information and Technology, Atomistic Modeling Center (AMC) · Munich Data Science Institute (MDSI), Technical University of Munich, Hans-Piloty-Straße 1, 85748 Garching, Germany · Institute of Structure of Matter – National Research Council Rome (ISM-CNR), via del Fosso del Cavaliere 100, Rome, 00133, RM,Italy +2
PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization
Oct 19, 2026Xinyi Chen, Jiahuan Pang, Titus Chima +2Molecular Property Prediction Self-Supervised Pre-Training
Global College, Shanghai Jiao Tong University, Shanghai, China. · Digital Innovation Research Center, Duke Kunshan University, Kunshan, Jiangsu, China. · Zu Chongzhi Center, Duke Kunshan University, Kunshan, Jiangsu, China. +1
EnSol: an environment-aware graph neural network for molecular solubility prediction
Sep 17, 2026Thao Nguyen, Saman Shafaei, Zhengyi Zhang +2Molecular Property Prediction Graph Neural Networks
Siebel School of Computing and Data Science, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA. · Carl R. Woese Institute for Genomic Biology, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA. · NSF Molecule Maker Lab Institute, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA. +2

Jun 11, 2026J/K move · Enter open · S save

A green solvent screening tool for emerging materials via uncertainty aware, transformer enhanced transfer learning

Ioannis Kouroudis, Simon Ternes, Zhaosu Gu, Gohar Ali Siddiqui+4

Chair of Simulation of Nanosystems for Energy Conversion, Department of Electrical Engineering, TUM School of Computation, Information and Technology, Atomistic Modeling Center (AMC) · Munich Data Science Institute (MDSI), Technical University of Munich, Hans-Piloty-Straße 1, 85748 Garching, Germany · Institute of Structure of Matter – National Research Council Rome (ISM-CNR), via del Fosso del Cavaliere 100, Rome, 00133, RM,Italy+2

Molecular Property Prediction

Open paper AskPDF arXiv

Accurate prediction of solubility remains a central challenge across materials science and sustainable chemistry. In particular due to emerging technologies like organic and hybrid photovoltaics, batteries, and catalysis, solvent usage is expected to increase significantly within the coming years. Therefore, substituting solvents with greener alternatives is vital. This is where machine learning can have substantial impact. However, the limited data on critical parameters of solubility significantly constraints machine learning efficacy. In this work, we transfer a pre-trained foundational model on QM9 targets to our application with minimal data requirements. Additionally, the pipeline integrates uncertainty quantification, allowing the user to gauge the confidence of the predictions. As baseline, we succeed in predicting the Hansen solubility parameters and Dielectric Constant for which extensive databases exist. Importantly, we achieve high model performance on additional targets, such as Gutmann Donor and Acceptor numbers, where the available data is extremely limited. Overall, we augment data on solubility descriptors by orders of magnitude with high quality predictions. For effective dissemination, we deploy easy-to-use, easily integrateable with high throughput labs, customizable tool for ranking and screening possible solvent substitutes. Finally, we rediscovered known green solvent alternatives and proposed new candidates proving its relevance for finding eco-friendly solvents.

Abstract

Materials Property Prediction

Explore similar work

A green solvent screening tool for emerging materials via uncertainty aware, transformer enhanced transfer learning

PRISM: Prior Relational Information for Self-supervised Modeling to Enhance Solubility OOD Generalization

EnSol: an environment-aware graph neural network for molecular solubility prediction