Bring in a table
On Setup, drop a CSV or Excel file, or paste a sheet. One of the two is enough. Each row is one formula you have already made and measured. Your table stays in this live session only — it is not saved to a database.
Product
Formulation Advisor
Optimum formulations with fewer experiments
What if an intelligent advisor could read your experimental data from this project, or from a past one, and recommend the next experiment to try?
As a product or process developer, you have a limited budget and a limited amount of time to develop a new product or process. You don't have the luxury of doing a full design of experiments, so you rely on your intuition and your knowledge to design the next trial and find the right composition or process settings.
You probably think that an AI needs tons of data before it can help. That is still true if you want a generic mathematical model. But when you are developing a specific product, you are fine-tuning. Bayesian optimisation looks at this small dataset that you already have (be it from this project or a prior similar project), and at the parameters you are tuning, and suggests a suitable next experiment to try.
That is what Formulation Advisor does. Advise you on your next experiment.
How to use
A short loop: load history, set one goal, compute a recipe, log the lab, repeat.
Map the columns
Mark what you can change: ingredients or process settings. For each one, say whether it is a number, a whole number, or a named option such as oil type. Named options can be unordered, or ordered such as low, medium, high. Lab measurements are things like viscosity or pH. Sensory scores are panel or consumer ratings. Each column goes in only one place.
Set one primary goal
Maximise, Minimise, or Target Spec on exactly one output. Floor, Ceiling, and Target Range on other outputs become constraints. Ignore means track only.
Compute
Compute next recipe opens Results: the next formula, a predicted value, and whether that suggestion stays near what you already measured or tries somewhere new. Balanced, Exploit, and Explore are set on Setup.
Log the lab, then compute again
Make the suggested formula, type the measured outputs on Results, and save the row. Compute again. The new row is highlighted in the table and the next suggestion updates. A bar chart shows each try. Scan the code on Results to keep going on a phone that is on the same Wi-Fi.
Inspect, don’t assume
Explore shows how the predicted result changes when you move one input, holding the others at the suggested recipe. Advanced is for experts. Leave those defaults unless you know why you are changing them. A change there applies only after you confirm it and compute again.
Formulation data is not saved to a product database. You can keep a Setup draft in this browser only (table + mappings + goals) so reload or idle pause does not lose your sheet. Closing clears the live session; use Clear browser draft to remove the local copy.
Read the full statement
What we process
Tables you upload or paste, tier mappings, goals, and lab rows you log are used only to fit models and suggest the next experiment in Formulation Advisor.
Where it lives
Live work is held in the memory of this application session (the server process backing the open pages). A phone that joins from Results shares that same memory. It is not stored in a product database on our side.
Browser draft (optional)
To survive reload or idle disconnect, this browser may keep your table, column mapping, goals, suggestion settings, and the last suggested experiment (so you can still run it in the lab and log the result). That copy stays on your device only. Use Clear browser draft on Setup to remove it. The fitted model itself is not stored — Explore predictions may need Compute again.
What is not kept
There is no account history vault and no durable experiment archive on our servers. When the live session ends — idle disconnect, you leave the page and any phone that joined, or the service restarts — in-memory data is gone unless you still have a browser draft.
FAQ
Trial and error, DoE, messy Excel data, and getting a next experiment with a small budget.
Is Bayesian optimisation better than trial and error for formulations?
Trial and error (and one-factor-at-a-time tweaks) rely on intuition and often miss interactions between ingredients or process settings. Bayesian optimisation learns from the experiments you already ran and suggests the next formulation or process trial that best balances improving the result with exploring uncertain regions — so you waste fewer lab runs chasing a local optimum.
How is this different from a design of experiments (DoE)?
Classical DoE plans a fixed set of trials to map effects across the design space. That is powerful, but the run count grows quickly with more variables and the plan usually does not adapt after each result. Formulation Advisor is for when you cannot afford a full DoE: it uses the small table you already have and recommends the next experiment iteratively. Many teams still start with a short screen, then switch to this loop for the remaining budget.
My lab data is messy — can I still use AI?
Yes. Real formulation sheets are rarely clean. You do not need a perfect database or thousands of rows. Bring a header plus measured outcomes, mark what you can change versus what you measure, and keep enough complete rows to learn from. The app accepts CSV or Excel (up to 10 columns and the last 100 data rows), coerces numeric fields, treats named options as categoricals, and drops incomplete rows for the fit. Clean only the columns that matter — do not wait for perfect data.
Can I get a next experiment suggestion from a small Excel or CSV table?
That is the main use case. Upload or paste your historical trials, set one primary goal (maximise, minimise, or hit a target), and compute a suggested next recipe. Log the lab result, then compute again. The model updates after each run instead of asking you to pre-plan every trial up front.
How many experiments do I need if I cannot afford a full DoE?
Bayesian optimisation is built for limited budgets and small datasets. A handful of past trials is enough to start; more complete rows usually improve the suggestion. If data is very thin, the tool explores more; as you log results it can shift towards exploiting the best predicted region for your last affordable batch.
What formulation inputs can I optimise?
Continuous amounts, integer settings, and categorical choices such as oil type or grade. Map ingredients and process knobs as what you change, lab measurements (viscosity, pH, …) and sensory scores as what you measure, then pick one primary goal.
Is my formulation data stored or shared?
Nothing is written to a product database. Live work stays in session memory. Optionally, a browser draft keeps your table, mappings, goals, and the last suggested experiment on this device so a lab step-away or idle pause does not break the loop — clear it anytime on Setup. A phone on the same Wi-Fi can join the live session from Results.
Is Formulation Advisor free to try?
Yes. Load a built-in example or your own table and run suggestions in the browser, subject to a fair daily trial limit per browser. Built by Decodexis for product and process developers who need better formulations with fewer experiments.