A quantum-transformer hybrid architecture for polymer property prediction: Addressing data sparsity issues

Computational Materials Science, Год журнала: 2025, Номер 256, С. 113950 - 113950

Опубликована: Май 16, 2025

Язык: Английский

---

Data‐Driven Design of High‐Performance Polyimides With Enhanced Heat Resistance and Dielectric Properties

Advanced Functional Materials, Год журнала: 2025, Номер unknown

Опубликована: Май 16, 2025

Abstract The evolution of electronic technology, such as high‐speed, high‐frequency, and high‐density integrated circuits, imposes higher performance requirements on advanced functional materials like polyimides. However, the prolonged development cycle linked with conventional trial‐and‐error methods results in a noticeable gap between material research its practical application. Here, genome approach is proposed to accelerate discovery polyimides exhibiting exceptional dielectric properties under elevated temperatures high frequencies. To address scarcity data, theoretical high‐frequency are derived by employing Havriliak‐Negami relaxation model complement experimental data. With augmented data polyimides, multi‐task learning hierarchical neural networks for glass transition temperature. Structural design via genetic algorithms implemented engineer polyimide structures enhanced properties. Several comprehensive generated, validation conducted. Shapley additive explanations analysis reveals crucial structural elements influencing performance. framework established this work can guide other polymeric materials.

Язык: Английский

---

Is BigSMILES the Friend of Polymer Machine Learning?

Опубликована: Авг. 20, 2024

The inherent randomness of polymers has long posed challenges for their representation learning in polymer machine (ML). Simplified Molecular-Input Line-Entry System (SMILES) notation, which excelled small molecule research, unfortunately, struggles to flexibly capture the complexity structures, such as random block copolymers. Recently, BigSMILES and its extensions have paved way more accurate descriptions structures. However, whether outperforms SMILES ML workflows yet be systematically explored demonstrated. To fill this scientific gap, we conducted extensive experiments investigating question, encompassing a variety property prediction inverse design tasks based on both image text inputs. Our findings reveal that 11 involving homopolymer systems, BigSMILES-based exhibit performance comparable or even exceeding SMILES, underscoring utility representing Furthermore, offers compact textual compared significantly reducing computational cost model training, particularly large language models. Through these comprehensive experiments, demonstrate can achieve par with while also facilitating faster training energy consumption, could substantial impact wide range future, including (and classification) generation across various types.

Язык: Английский

---

Toward Automated Simulation Research Workflow through LLM Prompt Engineering Design

Journal of Chemical Information and Modeling, Год журнала: 2024, Номер 65(1), С. 114 - 124

Опубликована: Дек. 30, 2024

The advent of Large Language Models (LLMs) has created new opportunities for the automation scientific research spanning both experimental processes and computational simulations. This study explores feasibility constructing an autonomous simulation agent (ASA) powered by LLMs through prompt engineering automated program design to automate entire process according a human-provided plan. includes design, remote upload execution, data analysis, report compilation. Using well-studied problem polymer chain conformations as test case, we assessed long-task completion reliability ASAs different LLMs, including GPT-4o, Claude-3.5, etc. Our findings revealed that ASA-GPT-4o achieved near-flawless execution on designated missions, underscoring potential methods like ASA achieve in enhance efficiency. outlined can be iteratively performed up 20 cycles without human intervention, illustrating workflow automation. Additionally, discussed intrinsic traits managing extensive tasks, focusing self-validation mechanisms, balance between local attention global oversight.

Язык: Английский

---

Generative Foundation Model for On-demand Reverse Polymer Design

Опубликована: Май 15, 2024

Forward screening and reverse design of drug molecules, inorganic polymers with better properties are crucial engines for shortening the laboratory-to-market cycle. Particularly, due to lack large-scale datasets, polymer discovery based on materials informatics is more formidable. Despite this, scientists have developed a series machine learning models structure-property relationships using only small thereby driving forward process polymers. However, success this paradigm ultimately hinges capacity candidate pool, while exhaustively enumerating all structures through human imagination challenging. Therefore, achieving on-demand crucial. In work, we curate dataset containing nearly one million polymeric pairs expert intuition. Using dataset, propose generative pre-trained model generation large language model. The produce 99.27\% chemical validity in top-1 mode (approximately 200k generated polymers), marking highest reported rate among models. addition, average $R^2$ between molecules their expected values across 15 predefined 0.96. To further assess model's performance generating additional user-defined downstream tasks, conduct fine-tuning experiments three publicly available datasets semi-template template-free paradigm. Through these extensive experiments, demonstrate that our fine-tuned capable specified properties, whether (semi-)template or challenging scenarios.

Язык: Английский

---

On-Demand Reverse Design of Polymers with PolyTAO

Опубликована: Июнь 6, 2024

The forward screening and reverse design of drug molecules, inorganic polymers with enhanced properties are vital for accelerating the transition from laboratory research to market application. Specifically, due scarcity large-scale datasets, discovery via materials informatics is particularly challenging. Nonetheless, scientists have developed various machine learning models polymer structure-property relationships using only small thereby advancing process polymers. However, success this approach ultimately depends on diversity candidate pool, exhaustively enumerating all possible structures through human imagination impractical. Consequently, achieving on-demand essential. In work, we curate an immense dataset containing nearly one million polymeric pairs based expert knowledge. Leveraging dataset, propose a Transformer-Assisted Oriented pretrained model generation (PolyTAO). This produces 99.27% chemical validity in top-1 mode (approximately 200k generated polymers), representing highest reported rate among generative models. Additionally, average R2 between their expected values across 15 predefined 0.96. To further evaluate model's performance generating additional user-defined downstream tasks, conduct fine-tuning experiments three publicly available datasets both semi-template template-free paradigms. Through these extensive experiments, demonstrate that our its fine-tuned versions capable specified properties, whether or more challenging scenarios, showcasing potential as unified foundation generation.

Язык: Английский

---

Identifying urban land use through higher-order spatial interactions

International Journal of Digital Earth, Год журнала: 2024, Номер 17(1)

Опубликована: Окт. 9, 2024

Crowd flow connects various geographic spaces in cities, revealing inter-regional associations that are crucial for urban land – use identification. Existing research mainly focuses on binary connections between pairs of regions, overlooking among multiple regions. Addressing this gap, we propose a network model uses hypergraph neural to extract key features higher-order classification. Additionally, similarity enhancement module is incorporated augment the recognition capabilities model. Compared with graph networks, incorporating regions improves Metrics show decrease 0.4 L1 distance, 2.35 KL divergence, and 0.14 Chebyshev while cosine increased by 0.25, particularly areas high crowd mobility. The further refines ability capture regional similarities, effective large contiguous or extreme points interest distributions. degree mixing movement influences effectiveness recognizing use, noticeable negative positive impacts, respectively. This study provides methods insights utilization land-use identification studies flow.

Язык: Английский

---