July 2026 Changelog

Most compute platforms tell you what a run cost after it finishes. That's appropriate for accounting but useless for engineering. By the time the invoice explains your decision, you've already made it. PHIN strives to estimate the cost before you commit, enabling informed techno-economic decisions. This month that number got materially more accurate: instead of pricing every task as though it had to run from scratch, estimates now recognize work your team has already done, quoting it at zero credits and cascading that reuse down the graph, so a study built on shared upstream work prices honestly instead of double-counting. Improving the estimation accuracy allows you to make real decisions about what is most important to run — how wide to sweep, which study to prioritize, whether a design fits the budget you have left. Every completed run's actual cost now appears alongside the estimate you approved, so you can see how well the two track.

In addition, a new Raman spectrum task predicts a full spectrum directly from a crystal structure, and the mixture and polymer builders now take a target density instead of a simulation box you have to derive yourself. We're also opening early access to the PHIN experiment agent for selected teams: describe what you want to study in plain language and it builds the experiment, wiring tasks and validating the graph before anything runs. Before running anything, the agent presents the experiment and its estimated cost at an approval gate, works under a per-session credit budget you set, and investigates failures on its own when a run goes wrong. The pattern is consistent: PHIN is built so the decisions happen with real numbers in hand.

Major features

• Accurate estimates, not worst-case ones: Cost estimates now recognize tasks whose results your team has already computed and quote them at zero credits, with reuse cascading through dependent tasks, so the number you see before running reflects what will actually happen.
• Estimate and actual, side by side: Each run's real cost now appears as a column in the experiment runs list, next to the estimate you approved.
• Raman spectra from structure: Predict a full Raman spectrum directly from a crystal structure with a new built-in analysis task.
• PHIN experiment agent, early access: Describe a study in plain language and have the experiment built, wired, and validated for you, with an approval gate and a credit budget standing between the agent and your balance. Available to selected teams.

Minor features

• Optimize against elastic constants: Individual elastic constants and other pattern-based properties can now be selected as optimization targets.
• Build by target density: Mixture, surface-mixture, and bulk-polymer generators accept a target density and infer the simulation box for you, opening a path on the polymer builder that was previously unavailable.
• Finer control over mixture composition: Concentration tolerance and maximum concentration error are now configurable on the mixture fraction generator.

Fixes & polish

• Structure relaxations using pre-trained machine-learning potentials now converge reliably, with existing configurations updated for you automatically. Renaming a task updates everywhere immediately, task connections in the experiment canvas link correctly, and an example template reports its lattice output under the right name.