PHIN journal
Research, case studies, product updates, and news from the team digitizing materials engineering.
PHIN-atomic, our first atomistic foundation model, reports a calibrated error bar on every energy and force it predicts. A scientist states the error a decision can absorb, and the model returns only the predictions that meet it, removing the jagged performance of pre-trained MLIPs at a fraction of what ensembling costs.
Read article
We demonstrate a core feature of PhinOS, the ability to replace expensive tool calling agent workflows with static task graphs to reduce LLM costs by 80-90%. We demonstrate this on a literature extraction workflow: same models, papers, and target schema, to demonstrate an up to 10x cost savings.
Read article
Over three years we proved that uncertainty-aware models can replace physical experiments in batteries, catalysts, water processing, and mining. Running that at scale takes a different product, so we rebuilt ours. PhinOS, our second-generation operating system for materials development, is now open to a closed group of external users.
Read article
Techno-economic modeling is required for engineering but most platforms tell you what a run cost after it finishes. In July, PHIN revamped its cost estimation to be much more accurate, by recognizing work your team has already done. Plus Raman spectra from structure, density-based builders, and early access to the PHIN experiment agent.
Read article
Coarse-grained simulations usually shrink each molecule to a featureless point, discarding the shape and orientation that make a polar liquid behave the way it does. We give every molecule a shape and an orientation and learn the interactions directly from data, capturing the full structure of liquid water, including correlations no point-particle model can represent, at 7-27x the speed of all-atom simulation.
Read article
June deepened PHIN's strengths as a platform for making decisions. You can now steer an experiment toward several competing goals at once and read the trade-offs directly, and more of the properties you care about are available straight from built-in templates.
Read article
Industrial materials scientists feel a tension between what AI simulation models achieve in benchmarks and what they experience in their materials development workflows. Despite datasets increasing by multiple orders of magnitude, this reliability gap still exists and much like the historical trajectory of Computational Fluid Dynamics (CFD), widespread industry adoption will not occur until there are strict, trusted error bounds. Phin addresses this tension through uncertainty quantification that identifies the reliability boundary of AI models, enabling the digitization of materials engineering.
Read article
Our recent panel discussion on Uncertainty Quantification (UQ) in Machine Learning Interatomic Potentials (MLIPs) explored the critical bottleneck of model reliability. Featuring insights from Jenna Pope (Pacific Northwest National Laboratory) and Nicholas Wimer (National Laboratory of the Rockies), we examined why deep ensembles aren't a universal solution and how post-hoc methods like quantile regression offer a more scalable path toward trustworthy, autonomous materials development, showcasing how PHIN is on the frontier of UQ architectures of MLIPs.
Read article
PHIN OS now uses AI workflows to reliability transform human goals into complex workflows that can predict experimental measurements
Read article
We are excited to deploy our AI agents and simulation technology with PNNL, WashU, and Lectrolyst to reduce the time to develop new industrial catalysts from decades to months.
Read article
PHIN's new Virtual Scientists—AI agents in the PhinOS platform—eliminate the "expertise barrier" that has long gated advanced materials simulation. Combining deep domain knowledge in materials science with PHIN's Active Learning infrastructure orchestration, these agents offer immediate answers and execute virtual experiments via a simple, conversational interface. This hybrid UX, which we call an “Instrumented Conversational UI” makes computational materials discovery accessible to all materials engineers, not just computational experts.
Read article
Battery simulations have traditionally required significant expertise and manual oversight to define simulation workflows and ensure accuracy. We show how both problems are addressed with PHIN-OS and PHIN-atomic. We leverage these tools to simulate the formation of solid electrolyte interphases, which remains a grand challenge in battery research.
Read article
We use PHIN-OS to simulate the Horiuti–Polanyi mechanism for acetylene hydrogenation, a widely used process for purifying industrial hydrocarbon streams. The case study validates a reproducible digital workflow that can subsequently be used to rapidly develop novel, cost-effective catalysts.
Read article
Case study of the nudged elastic band (NEB) workflow in PHIN-OS to predict Li diffusion in an LFP cathode.
Read article
Here we introduce a new paradigm in PHIN-OS, composable design space search that enables using state of the art Bayesian Optimization (BO) into any atomic scale simulation workflow. The functionality reduces the cost of materials design by 97+% by efficiently navigating vast design spaces.
Read article
PHIN OS and PHIN Atomic engine provide a powerful platform for performing high-throughput phonon calculations. Its modular design, active-learning based engine, and efficient task orchestration make it possible to explore phonon properties across diverse material classes with accuracy and speed beyond what is possible with a DFT-only approach. The computed phonon band structures and DOS are in good agreement with DFT-based reports, demonstrating that PHIN Atomic and PHIN OS are powerful tools for reliable high-throughput phonon calculations. The results presented here capture the essential vibrational features of MAX phases, NiTi shape memory alloys, and Mg₃Bi₂ thermoelectric, underscoring the utility of PHIN OS for high-throughput materials discovery, characterization, and design.
Read article
By simulating silicon vacancy energy, surface energy, and melting temperature, we showcase the ability of MLIPs to simulate real properties relevant to semiconductor development. We show that fine-tuning in PHIN-atomic is necessary to accurately simulate the properties of real materials and is a significant improvement over pretrained models.
Read article
We introduce a new concept in machine learning interatomic potentials, triplet envelope functions, inspired by geometric screening functions from classical many-body potentials, that doubles training and inference speed, triples memory efficiency, and ensures energy conservation.
Read article
Over the past year, we have reduced the data requirements 99% and the time by 97%
Read article
Cropping allows us to generate training data from complex simulations to ensure the accuracy of 1000+ atom simulations...
Read article
Understanding and analyzing a system’s response to various inputs is central to any scientific or engineering R&D. Designing a new steel...
Read article
Digitizing materials development requires materials models that can predict materials behavior accurately, quickly and cheaply. Digital...
Read article
Digitizing materials development requires the accurate prediction of materials properties at lower costs than an equivalent experiment . . .
Read article
Materials innovation is slow and complex, often taking decades. We started PHIN to bring trustworthy AI models to revolutionize materials development through a new paradigm of digital simulation.
Read article