Research

ID4 Research Highlights

We’re excited to share some of our most significant accomplishments:

– Kirigami Strata
– Crystal Generative Modeling
– Dynamic Metamaterials
– AI-ready Data: Lab Notebook Digitization
– Language Representations for Materials Exploration and Discovery

Kirigami Strata

ID4 researchers working on this project include faculty Sigrid Adriaenssens and Ryan P. Adams, as well as graduate students Isabel Moreira de Oliveira and Rafael Pastrana, and postdoc Cindy Zhang (left to right). External collaborators include Emily Baker of the University of Arkansas, Vittorio Paris of the University of Bergamo, and Taramelli Srl with its partner Carpenteria Bonatese Srl (not shown).

Kirigami Strata: Layers of Symmetry and Form showcases how mathematical transformations, defined by symmetry groups, achieve tailored patterns for kirigami space frames. Each pattern endows the space frame with unique, multifaceted attributes: remarkable mechanical performance, distinctive architectural expressions, and a mesmerizing interplay of light and shadow. The kirigami patterns are generated by a neural network, an artificial intelligence (AI) model that learns from data to optimize the performance of building components. This AI model serves as a creative tool, allowing designers to use text and image prompts to guide the design of these cut patterns. The synthesis of AI and mechanics supports designers in the regeneration of built environments where creativity, performance, and low-waste fabrication coexist. With this integrative framework, designers can create easy-to-build, lightweight and strong building components such as roofs, floors, and walls, as well as light-steering porosity for façade applications.

Kirigami Strata is led by Princeton University affiliates Isabel Moreira de Oliveira, Rafael Pastrana, Cindy Zhang, Sigrid Adriaenssens, and Ryan P. Adams, with the collaboration of Emily Baker of the University of Arkansas, Vittorio Paris of the University of Bergamo, Carpenteria Bonatese S.r.l., and Taramelli S.r.l.

The exhibit will be on display at the Palazzo Mora between May 10th and November 23rd, 2025 as part of the European Cultural Centre’s architecture exhibition Time Space Existence in Venice, Italy. For more information, see the exhibit website.

Return to top

Crystal Generative Modeling

ID4 researchers working on this project include graduate students Rees Chang, Angela Pak, Alex Guerra, and Nick Richardson (top row, left to right), postdoc Ni (Jenny) Zhan, and faculty Elif Erketin and Ryan P. Adams (bottom row, left to right). External collaborators include Sulin Liu (former ID4 graduate student), Ryan Marr, and Alex M. Ganose (not shown).

Inverse crystalline materials design is a grand challenge in materials science. Most crystals have atoms at high-symmetry subspaces of 3D Euclidean space; yet, most existing crystal generative models cannot place atoms in these positions with nonzero probability. We have developed two generative models that are the first to yield space group invariant likelihoods, naturally enforcing space group symmetry constraints and correctly assigning the same likelihood to atoms which are symmetrically equivalent. Wyckoff- and Asymmetric Unit-based Generative model (WyckoffAUGen) sequentially builds crystals with explicit autoregressive-like conditional likelihoods and hard space group constraints. Building on WyckoffAUGen and parallel work by the Adams and Bertoldi groups on equivariant flows for metamaterials, we developed SGEquiDiff, which introduces a space group equivariant diffusion model for crystals. SGEquiDiff achieves state-of-the-art performance on benchmark datasets as assessed by quantitative proxy metrics and quantum mechanical calculations, including outperforming existing models for generating unique and novel crystals that are predicted to be stable.

For more information, check out the papers for WyckoffAUGen and SGEquiDiff.

Return to top

Dynamic Metamaterials

Researchers working on this project include Sina Jafarzadeh of DTU, ID4 postdoc Giovanni Bordiga, ID4 graduate student Audrey Watkins, visiting undergraduate Yeqi Chu, and visiting professor Vincent Tournat of CNRS (left to right, across both photographs). Other ID4 researchers involved in this project are postdocs Eder Medina and Cyrill Bösch, and faculty Katia Bertoldi and Ryan P. Adams (not shown).

A metamaterial is an engineered material whose properties arise not just from its atomic/molecular composition, but from its designed internal structure, resulting in metamaterials exhibiting properties not available in natural materials. We have developed a framework to automate the design of flexible metamaterial structures that can execute desired nonlinear dynamic tasks. We focus primarily on designing flexible mechanical metamaterials constructed from rigid units with elastic couplings. Through a mix of advancements in automatic differentiation and gradient-based optimization, we demonstrate how the nonlinear dynamic response of these metamaterials can be tailored to execute complex tasks such as energy focusing, energy splitting, dynamic protection, nonlinear motion conversion, cloaking, and mechanical memory. In addition, our framework allows us to create architectures capable of seamlessly switching between different tasks and whose functionality can be reprogrammed on the fly.

For more information, check out our papers on reprogrammable nonlinear dynamic metamaterials, nonlinear mechanical metamaterial cloaks, and mechanical memory. Also check out the source code developed for the fully differentiable dynamic design framework for 2D flexible mechanical metamaterials, which is available on GitHub.

Return to top

AI-ready Data: Lab Notebook Digitization

This project is a collaboration between researchers at Drexel University and University of Central Florida. The researchers are: (top row) graduate student Joel Pepper and ID4 faculty Fernando Uribe-Romo, (middle row) REU student Zachary Siapno and ID4 faculty Jane Greenberg, and (bottom row) ID4 graduate students Jacob Furst and Xintong Zhao, REU student Elizabeth Jones, and faculty David Breen.

Collections of analog lab notebooks are an invaluable source of data about research conditions, steps, and outcomes, and in aggregate have the potential to provide new insights into the successes, failures and pedagogy of research laboratories. Unfortunately, these artifacts are increasingly at risk of being lost from the historical scientific record, given limited archiving and an absence of computational and AI readiness. A collaboration between the Greenberg and Uribe-Romo groups has allowed ID4 to advance an approach for extracting and structuring data from hand-written lab notebooks. This collaboration focused on transforming analog lab notebooks to AI-ready data by first making digital scans of hand-written lab notebooks, then leveraging OCR and object detection software to extract the contents, followed by conversion of the contents from plain text to a more machine readable format. Information from every stage of this process is collected into a database, allowing analysis of what was performed during laboratory experiments and ultimately to ML models for predicting successful synthesis conditions.

This work was first presented at the 2024 IEEE International Conference on Big Data (BigData); for more details check out the conference paper.

Return to top

Language Representations for Materials Exploration and Discovery

ID4 has multiple parallel efforts to use language representations to enable materials exploration and discovery. The above ID4 group photo (with a couple of key people added) identifies graduate students, postdocs, and faculty who are involved in these efforts.

Data-driven approaches to materials exploration and discovery are building momentum due to emerging advances in machine learning. Despite this, the ability to predict the properties of crystalline materials has remained limited. We have developed a number of new tools to predict the properties of crystalline materials, including thermoelectric (TE) materials and metal-organic frameworks (MOFs).

In one effort, which was a collaboration between the Ertekin and Toberer groups, we created a materials recommendation framework that utilizes language representations to explore the vast materials space. We used natural language embeddings from pretrained transformer‐based (BERT) models to represent inorganic crystals and then apply a “recall + ranking” funnel architecture: first retrieving candidates similar to a query material, then ranking them for desired material properties via multi-task learning. We applied this to TE material discovery, showing that the language‐derived representations both recall relevant candidates more effectively (including underexplored material classes) and produce competitive property predictions compared to specialized ML models. Experimental validation confirms that some of the recommended materials exhibit promising TE performance. For additional information, and to access our Material_Recommender code, see our paper on this approach.

In a parallel approach from the Dieng group, we created the LLM-prop framework by carefully fine-tuning the encoder part of a small version of the T5 model on text descriptions of crystal structures to learn crystal representations that are used to predict the physical and electronic properties of any crystal material (see details and access our LLM-Prop code in our paper on this approach). In a collaboration between the Dieng and Gómez-Gualdrón groups, we pre-trained LLM-prop on strain energy prediction and then fine-tuned on free energy prediction using MOFSeq, a novel sequence representation of MOFs developed in the Gómez-Gualdrón group. When tested on ~5,400 MOFs, this combined approach achieved 98% accuracy in identifying synthetically accessible structures and correctly selected the most stable polymorph in 78% of the polymorphic families observed. These results demonstrate the promise of our approach in accelerating the discovery of synthetically accessible MOFs. Check out our paper for more information.

In a final effort in this area, the Greenberg, Gómez-Gualdrón, and Uribe-Romo groups used LLMs to extract and codify synthesis procedures of MOFs from experimental literature. Specifically, we developed an end-to-end pipeline that combines literature matching, synthesis paragraph classification, and prompt-based entity and relation extraction using GPT-4. Guided by MOF experts, we designed a comprehensive and FAIR-compliant synthesis codification schema that captures synthesis actions, precursors, conditions, and their interrelations as a sequence-aware directed graph. Our model achieves high accuracy in synthesis paragraph classification (F1 score: 0.93) and in entity and relation extraction (F1 score: 0.96 and 0.94, respectively). This work enables large-scale, structured synthesis data collection and paves the way for AI-assisted synthesis prediction and knowledge discovery in materials science. For more information, check out our paper on this approach.

Return to top