Search for: All records

Abstract The prediction of crystal properties plays a crucial role in materials science and applications. Current methods for predicting crystal properties focus on modeling crystal structures using graph neural networks (GNNs). However, accurately modeling the complex interactions between atoms and molecules within a crystal remains a challenge. Surprisingly, predicting crystal properties from crystal text descriptions is understudied, despite the rich information and expressiveness that text data offer. In this paper, we develop and make public a benchmark dataset (TextEdge) that contains crystal text descriptions with their properties. We then propose LLM-Prop, a method that leverages the general-purpose learning capabilities of large language models (LLMs) to predict properties of crystals from their text descriptions. LLM-Prop outperforms the current state-of-the-art GNN-based methods by approximately 8% on predicting band gap, 3% on classifying whether the band gap is direct or indirect, and 65% on predicting unit cell volume, and yields comparable performance on predicting formation energy per atom, energy per atom, and energy above hull. LLM-Prop also outperforms the fine-tuned MatBERT, a domain-specific pre-trained BERT model, despite having 3 times fewer parameters. We further fine-tune the LLM-Prop model directly on CIF files and condensed structure information generated by Robocrystallographer and found that LLM-Prop fine-tuned on text descriptions provides a better performance on average. Our empirical results highlight the importance of having a natural language input to LLMs to accurately predict crystal properties and the current inability of GNNs to capture information pertaining to space group symmetry and Wyckoff sites for accurate crystal property prediction. 

Abstract The increasing integration of Visual Language Models (VLMs) into visualization systems demands a comprehensive understanding of their visual interpretation capabilities and constraints. While existing research has examined individual models, systematic comparisons of VLMs' visualization literacy remain unexplored. We bridge this gap through a rigorous, first‐of‐its‐kind evaluation of four leading VLMs (GPT‐4, Claude, Gemini, and Llama) using standardized assessments: the Visualization Literacy Assessment Test (VLAT) and Critical Thinking Assessment for Literacy in Visualizations (CALVI). Our methodology uniquely combines randomized trials with structured prompting techniques to control for order effects and response variability ‐ a critical consideration overlooked in many VLM evaluations. Our analysis reveals that while specific models demonstrate competence in basic chart interpretation (Claude achieving 67.9% accuracy on VLAT), all models exhibit substantial difficulties in identifying misleading visualization elements (maximum 30.0% accuracy on CALVI). We uncover distinct performance patterns: strong capabilities in interpreting conventional charts like line charts (76‐96% accuracy) and detecting hierarchical structures (80‐100% accuracy), but consistent difficulties with data‐dense visualizations involving multiple encodings (bubble charts: 18.6‐61.4%) and anomaly detection (25‐30% accuracy). Significantly, we observe distinct uncertainty management behavior across models, with Gemini displaying heightened caution (22.5% question omission) compared to others (7‐8%). These findings provide crucial insights for the visualization community by establishing reliable VLM evaluation benchmarks, identifying areas where current models fall short, and highlighting the need for targeted improvements in VLM architectures for visualization tasks. To promote reproducibility, encourage further research, and facilitate benchmarking of future VLMs, our complete evaluation framework, including code, prompts, and analysis scripts, is available athttps://github.com/washuvis/VisLit‐VLM‐Eval. 

Abstract Research shows that user traits can modulate the use of visualization systems and have a measurable influence on users' accuracy, speed, and attention when performing visual analysis. This highlights the importance of user‐adaptive visualization that can modify themselves to the characteristics and preferences of the user. However, there are very few such visualization systems, as creating them requires broad knowledge from various sub‐domains of the visualization community. A user‐adaptive system must consider which user traits they adapt to, their adaptation logic and the types of interventions they support. In this STAR, we survey a broad space of existing literature and consolidate them to structure the process of creating user‐adaptive visualizations into five components: Capture ⒶInputfrom the user and any relevant peripheral information. Perform computational ⒷUser Modellingwith this input to construct a ⒸUser Representation. Employ ⒹAdaptation Assignmentlogic to identify when and how to introduce ⒺInterventions. Our novel taxonomy provides a road map for work in this area, describing the rich space of current approaches and highlighting open areas for future work. 

Abstract The increasing integration of artificial intelligence (AI) in visual analytics (VA) tools raises vital questions about the behavior of users, their trust, and the potential of induced biases when provided with guidance during data exploration. We present an experiment where participants engaged in a visual data exploration task while receiving intelligent suggestions supplemented with four different transparency levels. We also modulated the difficulty of the task (easy or hard) to simulate a more tedious scenario for the analyst. Our results indicate that participants were more inclined to accept suggestions when completing a more difficult task despite theai's lower suggestion accuracy. Moreover, the levels of transparency tested in this study did not significantly affect suggestion usage or subjective trust ratings of the participants. Additionally, we observed that participants who utilized suggestions throughout the task explored a greater quantity and diversity of data points. We discuss these findings and the implications of this research for improving the design and effectiveness ofai‐guidedvatools. 

Abstract Photochemical C−C coupling reactions can be tailored to industrial chemical processes and preparations of pharmaceuticals. Recent approaches in this area are limited to using precious transition metal coordination complexes that facilitate light absorption and redox processes with benchtop chemicals. Herein, we propose a paradigm that involves all‐in‐one organo‐photo‐auxiliaries,thio‐heteroarenes, which exhibit unique photophysical properties. Thesethio‐heteroarenes were employed to prepare several all‐in‐one ionic photo‐salts from commercially available alkyl/benzyl and heterocyclic halides via aromaticity‐mediated nucleophilic substitution reactions. From the library of >30 salts, we performed on‐demand photochemical C−C coupling reactions to isolate numerous symmetrical and unsymmetrical diary/alkyl‐ethanes with yields up to 84% and mass balance as high as 96%. We also investigated the influence of structural features/properties on the outcomes of the photochemical C−C coupling reactions. The current photochemical C−C method was successful in the isolation of >30 photoproducts, including the natural product Brittonin A, a precursor of Imipramine, and derivatives of the bioactive Honokiol Analogues. Furthermore, transient absorption spectroscopy and time‐dependent density functional theory calculations were used to decipher the nature of light‐promoted electronic transitions. 

Abstract Data-driven approaches to materials exploration and discovery are building momentum due to emerging advances in machine learning. However, parsimonious representations of crystals for navigating the vast materials search space remain limited. To address this limitation, we introduce a materials discovery framework that utilizes natural language embeddings from language models as representations of compositional and structural features. The contextual knowledge encoded in these language representations conveys information about material properties and structures, enabling both similarity analysis to recall relevant candidates based on a query material and multi-task learning to share information across related properties. Applying this framework to thermoelectrics, we demonstrate diversified recommendations of prototype crystal structures and identify under-studied material spaces. Validation through first-principles calculations and experiments confirms the potential of the recommended materials as high-performance thermoelectrics. Language-based frameworks offer versatile and adaptable embedding structures for effective materials exploration and discovery, applicable across diverse material systems. 

Abstract Thermal energy management in metal-organic frameworks (MOFs) is an important, yet often neglected, challenge for many adsorption-based applications such as gas storage and separations. Despite its importance, there is insufficient understanding of the structure-property relationships governing thermal transport in MOFs. To provide a data-driven perspective into these relationships, here we perform large-scale computational screening of thermal conductivitykin MOFs, leveraging classical molecular dynamics simulations and 10,194 hypothetical MOFs created using the ToBaCCo 3.0 code. We found that high thermal conductivity in MOFs is favored by high densities (> 1.0 g cm⁻³), small pores (< 10 Å), and four-connected metal nodes. We also found that 36 MOFs exhibit ultra-low thermal conductivity (< 0.02 W m⁻¹ K⁻¹), which is primarily due to having extremely large pores (~65 Å). Furthermore, we discovered six hypothetical MOFs with very high thermal conductivity (> 10 W m⁻¹ K⁻¹), the structures of which we describe in additional detail. 

Abstract While machine learning has emerged in recent years as a useful tool for the rapid prediction of materials properties, generating sufficient data to reliably train models without overfitting is often impractical. Towards overcoming this limitation, we present a general framework for leveraging complementary information across different models and datasets for accurate prediction of data-scarce materials properties. Our approach, based on a machine learning paradigm called mixture of experts, outperforms pairwise transfer learning on 14 of 19 materials property regression tasks, performing comparably on four of the remaining five. The approach is interpretable, model-agnostic, and scalable to combining an arbitrary number of pre-trained models and datasets to any downstream property prediction task. We anticipate the performance of our framework will further improve as better model architectures, new pre-training tasks, and larger materials datasets are developed by the community. 

Abstract Materials with target nonlinear mechanical response can support the design of innovative soft robots, wearable devices, footwear, and energy‐absorbing systems, yet it is challenging to realize them. Here, mechanical metamaterials based on hinged quadrilaterals are used as a platform to realize target nonlinear mechanical responses. It is first shown that by changing the shape of the quadrilaterals, the amount of internal rotations induced by the applied compression can be tuned, and a wide range of mechanical responses is achieved. Next, a neural network is introduced that provides a computationally inexpensive relationship between the parameters describing the geometry and the corresponding stress–strain response. Finally, it is shown that by combining the neural network with an evolution strategy, one can efficiently identify geometries resulting in a wide range of target nonlinear mechanical responses and design optimized energy‐absorbing systems, soft robots, and morphing structures. 

Abstract Researchers collect large amounts of user interaction data with the goal of mapping user's workflows and behaviors to their high‐level motivations, intuitions, and goals. Although the visual analytics community has proposed numerous taxonomies to facilitate this mapping process, no formal methods exist for systematically applying these existing theories to user interaction logs. This paper seeks to bridge the gap between visualization task taxonomies and interaction log data by making the taxonomies more actionable for interaction log analysis. To achieve this, we leverage structural parallels between how people express themselves through interactions and language by reformulating existing theories asregular grammars.We represent interactions asterminalswithin a regular grammar, similar to the role of individual words in a language, and patterns of interactions ornon‐terminalsasregular expressionsover these terminals to capture common language patterns. To demonstrate our approach, we generate regular grammars for seven existing visualization taxonomies and develop code to apply them to three public interaction log datasets. In analyzing these regular grammars, we find that the taxonomies at the low‐level (i.e., terminals) show mixed results in expressing multiple interaction log datasets, and taxonomies at the high‐level (i.e., regular expressions) have limited expressiveness, due to primarily two challenges: inconsistencies in interaction log dataset granularity and structure, and under‐expressiveness of certain terminals. Based on our findings, we suggest new research directions for the visualization community to augment existing taxonomies, develop new ones, and build better interaction log recording processes to facilitate the data‐driven development of user behavior taxonomies.