📄 chemistry

Surrogate-Free Inverse Design of High-IP/High-EA Polyimide Constitutional Repeating Units with In-Loop Semiempirical Quantum Scoring

This study presents a surrogate-free inverse-design workflow combining recurrent-neural-network-guided Monte Carlo Tree Search with in-loop GFN2-xTB scoring to successfully generate a library of polyimide constitutional repeating units that simultaneously achieve deep ionization potentials and large electron affinities, overcoming the typical trade-off between these properties.

Original authors: Yingda Li, Rong Sun, Jibao Lu

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yingda Li, Rong Sun, Jibao Lu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect trying to design the ultimate "energy fortress." You need a material that is incredibly tough against losing electrons (like a vault that won't open) but also super eager to grab new ones (like a magnet that never lets go). In the world of chemistry, these two goals usually fight each other. It's like trying to build a house that is both a perfect fortress and a perfect trap; usually, making the trap stronger makes the fortress weaker.

For a long time, scientists trying to design these special materials (called polyimides) have relied on "crystal ball" guesses. They use computer programs trained on old data to predict how new molecules will behave. But here's the problem: when you try to design something totally new and extreme, those crystal balls often start hallucinating. They might tell you a molecule is a super-weapon when it's actually a dud, simply because the molecule is so different from anything they've seen before.

The New "No-Guessing" Approach
In this study, researchers Yingda Li, Rong Sun, and Jibao Lu decided to stop guessing. Instead of using a crystal ball, they built a digital workshop where every single new design is tested with a physics-based calculator called GFN2-xTB. Think of it like this: instead of asking a fortune teller if a new car engine will work, they actually built a tiny, fast prototype of the engine and revved it up to see what happened. They did this for every single candidate they created.

They used a smart computer brain (a Recurrent Neural Network) guided by a search strategy called Monte Carlo Tree Search. Imagine a giant maze where the computer is looking for the exit. At every turn, instead of guessing which path looks good, it runs a quick physics simulation to check the temperature, speed, and stability of the path before moving forward.

The Results: A Treasure Trove of New Designs
The team sent their digital explorer out to search through 360,846 potential molecular designs. Out of those, 179,316 were valid, stable structures that could actually exist.

The results were a massive success. In their original training data (the "old library" of known molecules), only 12.9% of the designs had the perfect combination of being tough against losing electrons and eager to grab them. But in their new, AI-generated library, 75.4% of the molecules hit the target! That is a sixfold improvement.

The "Goldilocks" Zone
The best designs they found are sitting in a special "Goldilocks" zone that nobody had really explored before for these types of materials.

Ionization Potential (IP): This measures how hard it is to steal an electron. The new designs have values between 7.97 and 9.01 eV. This is "deep," meaning they are very stable and hard to break.
Electron Affinity (EA): This measures how much the material wants to grab an electron. The top designs reached values between 2.85 and 3.99 eV.

To put this in perspective, the researchers compared their new designs to two other groups:

Old Polyimides: These had great electron-grabbing power but were too easy to break (shallow IP).
Standard n-type Semiconductors: These were great at grabbing electrons but were also too easy to break.

The new designs are unique because they manage to be both very stable and very good at grabbing electrons at the same time. One standout candidate (Candidate No. 40) grabbed electrons with an affinity of 3.99 eV while staying incredibly stable with an ionization potential of 8.49 eV.

How They Did It: The Secret Ingredients
The researchers broke down the winning molecules to see what made them special. They found that the best designs were built using specific "Lego blocks":

Imide Cores: The heart of the molecule was packed with carbonyl groups (like little magnets) that helped stabilize the electrons.
Fluorine and Cyanide: Adding fluorine atoms and cyano groups acted like a heavy anchor, pulling the energy levels down to make the molecule both stable and eager to accept electrons.
Avoiding Donors: They learned that adding "donor" parts (which like to give away electrons) was a bad idea because it made the molecule less stable.

What This Means (and What It Doesn't)
The paper suggests that these designs could be the building blocks for future electronics that work in the air without breaking down, or for high-temperature materials used in energy storage. The computer simulations show that these molecules are physically sound and follow the laws of quantum mechanics.

However, the authors are careful to note that these are still computer simulations. They haven't been built in a lab yet. The "synthetic accessibility" scores suggest they aren't impossibly hard to make, but the real test will be when chemists try to mix these ingredients in a beaker. Until then, these designs remain a powerful, physics-backed roadmap for what could be built, proving that if you stop guessing and start calculating, you can find hidden treasures in the chemical world that were previously invisible.

Technical Summary: Surrogate-Free Inverse Design of High-IP/High-EA Polyimide Constitutional Repeating Units

Problem Statement
The simultaneous optimization of deep ionization potentials (IP) and large electron affinities (EA) in organic materials is a significant challenge. Conventional strategies to stabilize the Lowest Unoccupied Molecular Orbital (LUMO) and increase EA (e.g., extending $\pi$ -conjugation) often raise the Highest Occupied Molecular Orbital (HOMO), thereby reducing IP. This trade-off is particularly acute for polyimide constitutional repeating units (CRUs), which are critical for high-temperature dielectrics and n-type organic electronics. While machine learning has been applied to polyimide property prediction, the deliberate inverse design of CRUs with IP and EA as independently co-optimized absolute energy-level targets has not been reported. Furthermore, standard generative design workflows relying on supervised surrogate predictors face a "surrogate labeling failure" when exploring out-of-distribution (OOD) regions of chemical space, where extrapolation errors can distort candidate ranking and lead to fictitious high-scoring structures.

Methodology
The authors developed a surrogate-free inverse-design workflow that couples a Recurrent Neural Network (RNN)-guided Monte Carlo Tree Search (MCTS) with in-loop semiempirical quantum-chemical scoring.

Generative Framework: The system utilizes the ChemTSv2 framework, where an RNN (pretrained on ~1.5 million polyimide-like CRUs) acts as a policy within an MCTS algorithm. The search explores the SMILES space through selection, expansion, rollout, and backpropagation steps.
In-Loop Evaluation: Unlike traditional workflows that use fast, trained surrogate models for property prediction, every generated candidate in the "rollout" phase undergoes an immediate GFN2-xTB (extended Tight-Binding) semiempirical calculation. This computes vertical IP and EA values directly from a quantum-mechanical Hamiltonian.
Reward Function: The reward $R$ is calculated as the sigmoid-normalized average of IP and EA: $R = \frac{1}{2}[\sigma(IP) + \sigma(EA)]$ . This formulation encourages the simultaneous maximization of both properties, penalizing trade-offs where one improves at the expense of the other.
Validation: A subset of generated candidates was validated using Density Functional Theory (DFT) at the B3LYP/aug-cc-pVTZ/COSMO(benzene) level to ensure rank consistency with the semiempirical screening.
Structural Analysis: Post-hoc analysis employed the BRICS algorithm to decompose successful CRUs into synthetically meaningful fragments, identifying statistical enrichment of specific motifs in the high-IP/high-EA region.

Key Contributions

Surrogate-Free Paradigm: The study demonstrates that in-loop semiempirical quantum evaluation (GFN2-xTB) can effectively replace supervised surrogates for inverse design, eliminating extrapolation errors in OOD chemical spaces while maintaining the throughput required for large-scale tree searches.
Novel Property Regime: The workflow successfully identified polyimide CRUs that occupy a distinct region of the IP–EA property space, characterized by deep IP values ( $>7.8$ eV) and large EA values ( $>1.5$ eV), a combination previously unreported for polyimide scaffolds.
Design Principles: The study provides interpretable, fragment-based design rules for achieving this property combination, moving beyond black-box generation to chemically actionable insights.

Results

Search Efficiency and Enrichment: The workflow explored 360,846 candidate SMILES strings, yielding 179,316 valid 3D geometries. Relative to the training set, the generated library showed a sixfold enrichment in the high-IP/high-EA region. Specifically, 75.4% of the generated candidates satisfied the criteria of IP > 7.8 eV and EA > 1.5 eV, compared to only 12.9% in the training set.
Validation Consistency: DFT benchmarking confirmed strong rank correlations between GFN2-xTB and higher-level DFT calculations for IP ( $R=0.93$ ), EA ( $R=0.98$ ), HOMO–LUMO gaps, optical gaps, and dipole moments. This validates the use of xTB for prioritizing candidates without systematic bias in ranking.
Top Candidates: The top-ranked CRUs achieved DFT-validated EA values of 2.85–3.99 eV while retaining deep IP values of 7.97–9.01 eV. Candidate No. 40 achieved the highest EA (3.99 eV) with a deep IP of 8.49 eV.
Literature Comparison: When compared to electrochemically modified polyimides and representative n-type organic semiconductors, the designed CRUs approach the electron-accepting strength of established n-type systems but maintain substantially deeper occupied levels (higher IP), effectively opening an underexplored region of molecular design space.
Structural Motifs: BRICS analysis identified that imide-rich cores (particularly fused multicarbonyl and cyano-substituted imides) are primary drivers for high EA. Polyfluoroaryl and compact fluorinated N-substituents act as cooperative motifs to enhance both IP and EA. Conversely, donor-rich bridging units were found to reduce IP and should be avoided for this specific objective.

Significance
The paper claims that this work establishes a robust, scalable, and interpretable strategy for the inverse design of polyimide CRUs targeting unusual electronic energy-level combinations. By avoiding surrogate predictors, the method ensures that property feedback remains physically grounded even in sparsely sampled regions of chemical space. The identified building blocks offer a potential route to ambient-stable n-type organic electronics (by stabilizing LUMO levels against oxidation) and high-temperature dielectric materials (via wide bandgaps and deep IP). The authors emphasize that while the results provide strong computational leads and design rules, experimental synthesis and polymer-level characterization are necessary next steps to validate these findings in real-world applications.

Technical Summary: Surrogate-Free Inverse Design of High-IP/High-EA Polyimide Constitutional Repeating Units

More like this