Agentic workflow enables the recovery of critical materials from complex feedstocks via selective precipitation

This paper presents a multi-agentic workflow that integrates AI agents with automated instruments to rapidly recover critical materials from complex real-world feedstocks via selective precipitation, reducing development timelines from years to days.

The Gist

Imagine you have a giant, messy bucket of soup. This isn't just any soup; it's a "complex feedstock" made of oil field water, crushed-up magnets, or other industrial waste. Inside this bucket, there are tiny, precious ingredients you desperately need—like rare metals used in your phone, electric cars, and wind turbines. But they are mixed with tons of useless (or even harmful) junk.

Traditionally, trying to fish out just the good ingredients from this soup is like trying to find a specific grain of sand on a beach while blindfolded. Scientists usually have to spend months or years testing thousands of chemical recipes, mixing things by hand, and hoping they get it right.

Enter CICERO: The AI Chef

This paper introduces a new system called CICERO (Computer Intelligence for Critical Elements Recovery and Optimization). Think of CICERO not as a single robot, but as a team of expert AI chefs working together in a high-tech kitchen.

Here is how this "Agentic Workflow" works, using simple analogies:

1. The Team of AI Agents

Instead of one super-computer trying to do everything, CICERO uses a team of specialized AI "agents" (digital assistants) that talk to each other:

• The Research Chef (Planning Agent): This agent reads thousands of cookbooks (scientific papers) and looks at the ingredients in your soup bucket. It says, "Okay, we have a lot of magnesium here. Let's try adding baking soda to pull it out." It creates a detailed recipe.
• The Economist (Viability Filter): Before the cooking starts, this agent checks the price of the ingredients. It asks, "Is it worth the cost to try this recipe?" If the answer is no, it stops the plan immediately, saving time and money.
• The Taste-Tester (Bayesian Optimization Agent): After the first batch of soup is made, this agent tastes it. If it's too salty, it doesn't just guess the next recipe; it uses math to figure out the exact tweak needed to make it perfect faster. It learns from every mistake.

2. The Automated Kitchen

Once the AI team agrees on a recipe, they don't ask a human to mix the chemicals. They send the instructions to a robotic arm (an Opentrons robot) in the lab.

• The robot acts like a super-fast, super-precise sous-chef. It can mix 96 different variations of the recipe at the same time (like a 96-well plate, which is just a tray with 96 tiny cups).
• It adds the exact drop of acid or base needed, stirs it, and spins it in a centrifuge (like a salad spinner) to separate the solids from the liquid.

3. The "Closed Loop" (The Magic Part)

This is where the "Agentic" part shines. In the old days, a scientist would mix the soup, wait for results, come back the next day, analyze the data, and then manually decide what to do next.

With CICERO, the loop is closed and automatic:

1. Plan: The AI Chef designs the experiment.
2. Do: The Robot executes it.
3. Analyze: A machine measures the results.
4. Refine: The AI Taste-Tester looks at the data and immediately designs the next experiment without human help.

It's like a video game where the character levels up automatically after every battle, getting smarter and stronger without the player needing to press a button.

The Results: What Did They Cook Up?

The team tested this system on three very different "soups":

1. Oil Field Water: They successfully pulled out pure Magnesium (used in alloys) with 99% purity in just one round of testing.
2. Crushed SmCo Magnets: They figured out how to separate Samarium (a rare earth metal) from Cobalt. The AI realized that by tweaking the pH just right, it could isolate the Samarium.
3. Crushed NdFeB Magnets: This was the hardest challenge. They had to separate Neodymium and Praseodymium (two very similar metals) from a massive amount of Iron. The AI figured out a two-step process: first, use oxalate to dump the iron, then use a specific pH tweak to separate the two rare metals.

The Big Win:
What used to take a human scientist months or years of trial and error, CICERO did in about three days.

Why This Matters

Think of the world's demand for critical materials (like those in your smartphone) as a hungry crowd at a buffet. The food is there, but it's buried under a mountain of trash.

• Old Way: Send one person to dig through the trash with a spoon. It takes forever, and they might miss the good stuff.
• CICERO Way: Send a team of AI chefs with a fleet of robotic arms. They scan the trash, figure out the best way to sift it, and pull out the gold in record time.

This paper proves that by letting AI agents plan, execute, and learn from experiments on their own, we can solve complex chemical problems much faster, cheaper, and more efficiently than ever before. It's not just about automation; it's about acceleration.

Technical Summary

1. Problem Statement

The recovery of critical materials (e.g., rare earth elements, cobalt, magnesium) from complex, real-world feedstocks is increasingly urgent due to surging demand for energy technologies.

• The Challenge: Traditional selective precipitation methods struggle with complex feedstocks (e.g., produced water, recycled magnets) because they contain multiple competing ions with similar physicochemical properties. This leads to impure products (inclusions/solid solutions) requiring further purification.
• The Bottleneck: Optimizing precipitation conditions across a vast parameter space (reagent types, concentrations, pH, sequence) using traditional trial-and-error or physics-based models is slow, often taking months or years. Existing AI tools are either too rigid (requiring heavy coding) or lack the specific domain knowledge to handle complex chemical workflows autonomously.

2. Methodology: The CICERO Workflow

The authors developed CICERO (Computer Intelligence for Critical Elements Recovery and Optimization), a multi-agentic workflow built on the SciLink platform. CICERO autonomously plans, executes, and optimizes separation campaigns using a closed-loop system.

A. System Architecture

The workflow integrates Large Language Models (LLMs) with automated laboratory instrumentation (Opentrons robots, ICP-MS). It operates through three coordinated AI agents:

1. Planning Agent:
   • Function: Ingests domain literature, technical reports, and experimental data via Retrieval-Augmented Generation (RAG).
   • Output: Generates testable hypotheses, detailed experimental protocols, and technoeconomic viability screenings (filtering out economically unviable paths before experiments begin).
   • Constraint: Restricted to using simple, commodity chemicals (e.g., NaOH, NaHCO₃, oxalic acid) to ensure scalability.
2. Scalarizer Agent (Data Processing):
   • Function: Converts raw instrument data (ICP-MS) into scalar descriptors suitable for mathematical optimization.
3. Bayesian Optimization (BO) Agent:
   • Function: Drives the closed-loop optimization. It uses Gaussian processes to model the parameter space and suggests the next set of experiments to maximize purity and yield, balancing exploration and exploitation.

B. Experimental Loop

1. Diagnose: Characterize feedstock chemistry via ICP-MS.
2. Evaluate & Hypothesize: The Planning Agent identifies target critical materials and generates a hypothesis for selective precipitation.
3. Experiment: The agent generates Python code to execute 96 experiments on an Opentrons liquid handling robot (96-well plates).
   • Human Role: Minimal. Humans move plates between the robot, centrifuge, and ICP-MS; they supervise code generation for minor bug fixes and set safety constraints (e.g., max reagent concentration).
4. Refine: Results are analyzed. If targets are not met, the BO Agent or Planning Agent iterates, suggesting new conditions for the next round.

3. Key Contributions

• Agentic AI for Chemistry: Demonstrated a fully functional "multi-agentic" system where LLMs autonomously plan, code, and refine chemical separation workflows without human intervention in the core scientific reasoning.
• Technoeconomic Filtering: Integrated economic viability screening directly into the hypothesis generation phase, ensuring the AI prioritizes processes that are not just chemically feasible but also economically scalable.
• Rapid Optimization: Reduced the timeline for developing separation protocols from months/years to days.
• Adaptability: The system successfully pivoted between fundamentally different chemical strategies for three distinct feedstocks without pre-programmed rules for each specific case.

4. Experimental Results

The CICERO workflow was tested on three diverse feedstocks:

A. Produced Water (Oklahoma)

• Target: Magnesium (Mg) recovery (1,840 ppm) amidst high Calcium (Ca) and Sodium (Na).
• Strategy: Selective precipitation of Mg(OH)₂ using NaOH.
• Outcome: Achieved 99.4% purity Mg(OH)₂ at 86% yield in a single iteration. The system successfully avoided co-precipitation of Ca, Sr, and Ba by tuning NaOH equivalents. Bulk scale-up (1000x volume) confirmed similar performance.

B. SmCo Permanent Magnets (Leached)

• Target: Samarium (Sm) recovery from a mix of Sm and Cobalt (Co).
• Strategy: Dual-reagent approach using NaHCO₃ and NaOH to alter Co speciation while precipitating Sm.
• Outcome:
  • Round 1: 85.0% Sm purity (Separation Factor ~22.8).
  • Round 2 (BO-optimized): Purity increased to 89–99% with >92% yield. The BO agent successfully narrowed the search space to optimal concentrations of NaHCO₃ and NaOH.

C. NdFeB Permanent Magnets (Leached)

• Target: Separation of Neodymium (Nd) and Praseodymium (Pr) from Iron (Fe) and Boron.
• Strategy:
  • Round 1: Used Sodium Oxalate (Na₂C₂O₄) to selectively precipitate Rare Earth Elements (REEs) over Iron. Achieved >99% purity REE salts with 93% yield and a separation factor of ~200 against Iron.
  • Round 2: Focused on separating adjacent elements Nd and Pr. The agent suggested pH adjustment with oxalate, achieving a Nd/Pr separation factor of 1.35 in a single round (notably difficult for adjacent lanthanides).
• Timeline: The entire optimization for NdFeB was completed autonomously in approximately three days.

5. Significance and Impact

• Scalability: By focusing on commodity chemicals and standard liquid handling, the workflow produces separation methods that are immediately scalable to industrial levels, unlike methods relying on bespoke membranes or ligands.
• Data Provenance: The system generates a reproducible knowledge base with full traceability from hypothesis to execution, addressing the "black box" nature of some AI applications.
• Future of Separation Science: This work proves that agentic workflows can handle the complexity of real-world waste streams, offering a pathway to rapidly develop feedstock-specific solutions for the circular economy. It shifts the paradigm from "human-guided AI" to "AI-guided human supervision," drastically accelerating materials recovery research.