Can AI Optimize Raw Material Processing? Or Just Help Us Understand It Better?
Mining gets most of the attention, but it’s what happens after you pull material from the ground that really determines whether it becomes something useful. Raw material processing is where chemistry, variability, and scale collide. It is where things can get very complicated very quickly.
Unlike mining, which plays out over decades and miles, processing happens in real time. Inputs shift by the hour, impurities creep up, equipment degrades, and small deviations in process control can ripple across a system and destroy yield, quality, or both.
That’s what makes this stage such an interesting target for AI. In theory, smarter tools could help stabilize processes, keep impurities in check, and guide flowsheet decisions based on shifting feedstock profiles. However, the reality is messier. Much of the relevant data doesn’t exist, or isn’t reliable, and the physical systems we’re working with weren’t built to accommodate algorithmic feedback loops.
This post looks at where AI is starting to make an impact, and where it still struggles, in the messy middle between resource and battery-grade material output.
What’s Working
AI is beginning to find real traction in areas where there’s sufficient data, real-time feedback, and a clear cost-benefit. In raw material processing, that typically means targeting yield, quality, and uptime.
1. Yield Maximization AI models can continuously adjust process parameters like temperature, residence time, and reagent dosing to push recovery rates higher without overstepping quality limits. Especially in multi-step processes like solvent extraction or crystallization, even small yield gains can have outsized economic value. These types of strategies are already being deployed in metals and chemical processing by companies like FLSmidth and Honeywell, and are beginning to be explored in lithium refining.
2. Real-Time Quality Control With sensors tracking lithium concentration, impurity levels (like magnesium or calcium), and physical properties, ML tools can detect deviations before they snowball. Combined with feedback loops, this lets operators keep output within spec and avoid costly reprocessing or process down time. Analogous systems are already used in flotation and comminution circuits with platforms like MineSense and FrothSense.
3. Process Flow Optimization This is less about real-time tweaks and more about designing the right flowsheet for a given feedstock. AI can help navigate tradeoffs in selectivity, reagent compatibility, and downstream integration, especially for complex brines or unconventional clay deposits. While still early, this area is attracting serious interest for decision support during piloting and scale-up.
4. Predictive Maintenance Chemical refinement can be especially aggressive on processing equipment. AI-powered maintenance models can spot early signs of trouble and reduce unplanned downtime, which is especially valuable in continuous or high-throughput systems. Tools developed in adjacent industries by firms like AspenTech, GE Digital, and ABB are beginning to influence thinking in the lithium space.
None of these applications are futuristic. They’re already being tested or deployed in pockets across the industry. However, they require a solid digital foundation, one that many plants still lack and may take time to employ.
What’s Missing
For all the promise, there are still big gaps when it comes to making AI broadly useful across the diverse and variable world of raw material processing.
1. Data Scarcity and Fragmentation It’s not just that data is limited. The data that does exist is fragmented across companies and formats. Each company guards its own historical process data, either to protect IP or to avoid training models that could benefit competitors. As a result, AI efforts are typically confined to narrow, proprietary datasets. That makes it much harder to build robust models or apply insights across different sites and systems.
2. Feedstock Variability No two brines, rocks, or clays are alike. This variability makes it hard to generalize models across sites. What works well for one feedstock can completely break down on another, especially in processes like DLE, where ion ratios, temperature, and fouling behavior can shift dramatically from one type of brine to another. It may turn out that each resource will require its own tailored model.
3. Black-Box Models and Lack of Domain Context Many AI tools are still black boxes. They might fit the data, but they don’t necessarily reflect chemical reality. This shortcoming makes operators hesitant to trust their outputs when a bad recommendation can damage equipment or send off-spec product downstream.
4. Missing Materials Data for AI-Driven Discovery Unlike cathode development or drug discovery, the field of extraction materials, adsorbents, solvents, membranes, isn’t backed by large, open datasets or supported by data from an academic community. This makes it hard to apply AI to design new materials for selective lithium (or any critical mineral) recovery or impurity rejection. Without high-quality, diverse data on how these materials behave across real-world conditions, model-driven discovery is mostly stuck at the starting line.
These gaps don’t mean AI has no place in processing. They just mean we need better data infrastructure, more collaborative experimentation, and more hybrid models that combine first-principles chemistry with machine learning.
What’s Next
The next wave of impact won’t come from retrofitting AI into broken systems, it will come from building smarter systems from the start. That means flowsheets designed with sensing, feedback, and optionality in mind. It also means investing in the boring stuff, such as data pipelines, rigorous calibration protocols, and human-in-the-loop engineering.
We’ll likely see:
- Hybrid models that combine physics-based logic with ML prediction
- AI-assisted flowsheet design tools during pilot development
- Digital twins that simulate process behavior under changing conditions
- AI-guided maintenance planning embedded into plant control systems
The most transformative potential may come from collaboration. Across the sector, we need better coordination between resource owners, operators, researchers, and technology developers to build shared datasets and open benchmarks. Without that, even the best models will remain stuck in the lab.
At EnergyX, we’ve built a platform that spans multiple extraction technologies, from membranes to sorbents to solvent-based systems, not because it’s convenient, but because it was necessary. Brines vary and requirements change. A single-technology will only get you so far. That diversity of tools gives us the flexibility to adapt and unlock new opportunities in the future. That same versatility puts us in a strong position to benefit from AI, both in accelerating our technology development and in moving faster toward commercialization.
If you’re working at the intersection of AI, process design, or materials science (especially in the lithium space), and want to explore what’s next together, we’d love to connect.
Progress in this space won’t come from any one company or breakthrough. It will take shared data, shared learning, and open-minded collaboration. Let’s build toward that future.
By: Dr. Nicholas Grundish