DFT Screening for Lithium Extraction Solvents

ISEF Category: Chemistry

Subcategory: Computational Chemistry  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Lithium powers phones, laptops, and electric cars, but getting it out of salty water is hard. Think of your project like sorting one tiny marble out of a jar full of almost identical beads. You can use computation to test many solvent ideas fast, then spend heavy DFT compute only where it matters most. That mix of chemistry and machine learning makes a strong research story.

What Is It?

Deep-eutectic solvents are liquid mixtures made from two or more components that lower each other's melting point. That can make them useful for pulling target ions, like lithium, out of a salty solution. Your job is to compare candidate solvent systems and predict which ones bind lithium more selectively than competing ions such as sodium or magnesium.

The core idea is a ranking problem. You start with a cheaper method, xTB, to get rough molecular properties fast. Then you send the most promising or most uncertain candidates to DFT, which stands for density functional theory, a more accurate but slower quantum chemistry method. A Gaussian-process surrogate acts like a smart shortcut. It learns from the results you already have and predicts which next calculations will teach you the most.

Why This Is a Good Topic

This is a strong science fair topic because it has a clear question, measurable outputs, and room for real modeling skill. You can compare binding energies, selectivity trends, and model error, all without needing a wet lab. The project connects to lithium mining, brine treatment, and greener separation methods, so the chemistry has real-world stakes. You can also learn how modern research teams save compute with active learning and surrogate models.

Research Questions

• How does the choice of deep-eutectic solvent components change predicted lithium binding energy?
• What is the effect of the initial xTB screening set on how many DFT calculations the surrogate model needs?
• Does a Gaussian-process surrogate reduce prediction error more than a simple linear baseline for lithium selectivity ranking?
• To what extent do candidate solvents distinguish lithium from sodium and magnesium in simulated brine conditions?
• Which molecular descriptors best predict lithium binding strength in the solvent screen?
• How does adding uncertainty-driven sampling change the final top-ranked solvent list?

Basic Materials

• A laptop or desktop computer with enough memory for quantum chemistry software.
• Access to a Linux environment or remote compute server.
• A spreadsheet program for tracking candidates and results.
• Python installed with NumPy, pandas, scikit-learn, and matplotlib.
• Open-source xTB software.
• Access to a DFT package approved by your lab, such as ORCA or Psi4.
• A set of candidate deep-eutectic solvent structures from literature review.
• Brine composition data from NOAA, USGS, or peer-reviewed papers.
• A reference manager for keeping track of papers and methods.

Advanced Materials

• High-performance computing access with job scheduling.
• ORCA, Psi4, or another DFT package with geometry optimization and single-point energy tools.
• xTB or similar semiempirical screening software.
• Python with ASE, RDKit, NumPy, pandas, scikit-learn, scipy, and matplotlib.
• Gaussian-process tooling such as scikit-learn or GPyTorch.
• Molecular visualization software such as Avogadro or VMD.
• A database of literature deep-eutectic solvent candidates.
• File storage for many geometry and output files.
• Optional scripting tools for automated workflow management.

Software & Tools

• Python: Runs data cleaning, feature generation, model training, and plotting for your screening workflow.
• scikit-learn: Builds Gaussian-process models and baseline regressors for active-learning comparisons.
• matplotlib: Makes learning curves, error plots, and solvent-ranking charts.
• ASE: Helps you manage molecular structures and automate quantum chemistry inputs and outputs.
• Avogadro: Lets you inspect and edit molecular geometries before running calculations.

Experiment Steps

1. Define the exact selectivity problem you want to solve, such as lithium over sodium or magnesium, and choose a small family of deep-eutectic solvents to compare.
2. Build a candidate list from review papers, then decide which molecular features or descriptors you will track for each solvent.
3. Run a cheap first-pass screen with xTB, then use those results to decide which systems deserve higher-level DFT follow-up.
4. Train a Gaussian-process surrogate on the early results, and plan how uncertainty will guide the next round of calculations.
5. Set controls that compare active learning against a random-selection baseline so you can prove the shortcut actually helps.
6. Predefine your scoring method for ranking solvents, checking error, and reporting whether your model changed the final top hits.

Common Pitfalls

• Mixing solvent families that are too different, which makes your model compare unrelated chemistry instead of a fair candidate set.
• Using unstable molecular geometries, which can cause xTB and DFT to disagree for numerical reasons instead of chemical ones.
• Training the surrogate on too few examples, which makes Gaussian-process uncertainty look confident even when the ranking is weak.
• Comparing binding energies without a consistent reference state, which can turn the selectivity metric into noise.
• Forgetting a random-sampling baseline, which makes it impossible to tell whether active learning actually saved calculations.

What Makes This Competitive

A stronger version of this project would test a clear active-learning strategy against at least one baseline and report both prediction error and compute saved. You can raise the level by using multiple ion competitors, careful uncertainty analysis, and a holdout set of solvents that the model never sees during training. Strong entries also explain why the chosen descriptors matter chemically, not just statistically. If you can connect the ranking to real brine chemistry, your project feels much closer to research.

Project Variations

• Screen choline-based deep-eutectic solvents instead of a broad solvent family, and compare lithium selectivity against magnesium-rich brine.
• Swap binding energy for solvation free-energy proxies, and test whether the ranking changes under a different chemical metric.
• Compare Gaussian-process active learning with random forest or neural-network uncertainty methods to see which saves the most DFT runs.

Learn More

• PubMed: Search for review articles on deep-eutectic solvents, lithium extraction, and brine separation chemistry.
• NIH PubChem: Look up molecular structures, names, and properties for solvent components.
• USGS Lithium resources: Find background on lithium sources, extraction, and supply context.
• NOAA data portals: Find brine, water chemistry, and environmental context data for salt-rich systems.
• MIT OpenCourseWare, Computational Chemistry or Physical Chemistry courses: Review quantum chemistry, molecular orbitals, and model-building basics.
• Journal articles in Green Chemistry and Journal of Chemical Theory and Computation: Search for recent papers on deep-eutectic solvents, DFT screening, and surrogate models.