Supercharging Battery Research: How AI is Predicting Lithium’s Next Move!

Hey there, science enthusiasts and tech lovers! Ever wonder what makes your phone last (or not last!) all day? A huge part of that magic comes down to tiny little things called lithium ions and how they zip around inside your battery. Finding materials where these ions can move super-fast is like striking gold for better batteries, zippier sensors, and all sorts of cool electrochemical gadgets. But here’s the rub: figuring this out traditionally involves some seriously heavy-duty calculations that can take ages and cost a pretty penny. I’m talking about density functional theory (DFT) – powerful, sure, but not exactly speedy for sifting through thousands of potential materials. So, what’s a material scientist to do? Well, we’re turning to our clever digital pals: machine learning (ML)!

The Old Grind: Why We Need a Faster Lane

Imagine you’re trying to design the next-gen solid-state battery. You need to understand how easily lithium ions can hop, skip, and jump through a material. This “ease of movement” is often boiled down to something called the migration barrier (Em) – basically, the energy hill an ion has to climb to get from one cozy spot to another in a crystal lattice. For decades, DFT has been our go-to for calculating these barriers. It’s pretty good, but it has its quirks.

For instance, the nudged elastic band (NEB) method, a common DFT sidekick, can be slow if you don’t give it a good starting guess for the ion’s path. Sometimes, it even gets lost and finds the wrong path altogether! Plus, it tends to underestimate these energy barriers by a bit because it doesn’t fully capture how the whole crystal lattice jiggles around at real-world temperatures. Another approach, ab initio molecular dynamics (AIMD), does consider these jiggles and often gives results closer to what we see in experiments. But, oh boy, is it computationally hungry! We’re talking simulations of just a few hundred atoms for tiny fractions of a second. This can lead to inaccuracies because of the small system size and not seeing enough ion hops to get a clear picture.

Given these hurdles, you can see why we’ve been itching for faster, more efficient ways to screen materials. We want to move from a trial-and-error slog to a high-throughput (HT) discovery pipeline. Some clever folks have come up with nifty tricks, like “pinball” models where the host lattice is frozen (saves computation time!) or using DFT charge density to find optimal migration paths. There’s also the Bond Valence Site Energy (BVSE) method, an empirical force field that’s pretty good at estimating a “percolation barrier” – the minimum energy for an ion to weave its way through the material. It correlates well enough with DFT and experiments to be a handy rapid screening tool.

Enter Machine Learning: The Game Changer?

With a veritable explosion of materials data, both from experiments and theory, the demand for speedy calculations of ionic transport is through the roof. This is where machine learning struts onto the stage. The idea is simple: if we can train ML models on existing data, maybe they can learn to predict these crucial transport properties much, much faster than traditional methods. We’ve already seen various data-driven solutions pop up, using everything from crystal structure descriptors to deep neural networks and classical ML models. Neural networks usually need a mountain of data to learn effectively, while classical ML models can get by with less but often require us to be really smart about crafting the features (the input information) we feed them.

Now, developing ML models to predict these energy barriers has been a hot topic, but there have been a few catches. Firstly, a lot of the datasets used for training have come from the faster, but less fundamentally accurate, BVSE method. DFT-derived data? A bit scarce. Secondly, many studies have zoomed in on specific types of crystal structures (like NASICONs or argyrodites) or a narrow range of chemical compositions. And thirdly, the datasets have often been on the smaller side. So, the big question has been hanging in the air: can we actually build an ML model that predicts Li-ion migration barriers for almost any material with DFT-level accuracy? That’s the dream!

Then came another breakthrough: interatomic ML potentials (MLIPs). These are reported to hit DFT-level accuracy for migration barriers and conductivity at a fraction of the cost. Super promising! But, typically, these potentials are custom-made for specific chemical systems, not ideal for exploring a vast, diverse chemical space in a high-throughput way. The latest buzz is around universal machine learning interatomic potentials (uMLIPs). Think of models like M3GNet, CHGNet, SevenNet, and MACE. These are often graph neural networks (GNNs) trained on massive datasets like the Materials Project, covering a wide range of chemistries. The hope is that these uMLIPs can be used for all sorts of downstream tasks, including modeling diffusion, and really kickstart the discovery of new materials. But, even with their application to Li-ion diffusion, we haven’t really had a solid benchmark to see how robust and accurate they are across the board.

Our Mission: Benchmarking the Bots!

This is where our work comes in! We wanted to really put these ML models to the test for HT screening of Li-ion conductors. To do this, we first had to build a solid foundation: a comprehensive dataset we’ve dubbed LiTraj. It’s a beast, containing:

• BVEL13k: ~13,000 percolation barriers calculated with the BVSE method.
• nebBVSE122k: A whopping ~122,000 migration barriers from BVSE-NEB calculations.
• nebDFT2k: ~1,700 migration trajectories and barriers crunched with DFT-NEB (using BVSE-NEB paths as a starting point).
• MPLiTrj: Data from DFT-NEB optimization of nearly 2,700 migration trajectories.

With LiTraj in hand, we set out to see if classical ML models and GNNs could tell “fast” ionic conductors from “poor” ones using these percolation and migration barriers. And, super importantly, we wanted to see how well those fancy uMLIPs could identify the best paths for Li-ions to travel and how accurate their energy predictions were. The exciting part? Fine-tuned uMLIPs are getting tantalizingly close to DFT accuracy, which could massively speed up our hunt for new ionic conductors!

Round 1: Percolation Barriers – Classical ML vs. GNNs

First up, we used our BVEL13k dataset (12,807 Li-ion 1D, 2D, and 3D percolation barriers) to benchmark classical ML models (like random forests, XGBRF, KNN) and GNNs (Nequip, Allegro, M3GNet). We fed the classical models features derived from Voronoi tessellation of the crystal structures – basically, carving up the structure into cells around each atom and looking at their geometric and elemental properties. We also threw in a general-purpose featurization scheme from Matminer for comparison.

The results? For 1D and 2D percolation, our custom features generally did better for most classical models. For 3D barriers, Matminer’s features with an XGBRF model took the crown. But here’s the kicker: the GNN models generally outperformed the classical gang, especially for 2D and 3D percolation barriers. Interestingly, M3GNet, with fewer parameters, did just as well as a much larger Nequip model for 1D barriers. It seems GNNs are pretty good at figuring out these structure-property relationships on their own! We even found that for predicting 1D percolation barriers, things like the fraction of covalent free space and the volume around Li sites were super important features. For a KNN model, around 30-35 features seemed to be the sweet spot for good accuracy without too much computational fuss.

Round 2: Migration Barriers – Digging Deeper

Next, we tackled the Li-ion migration barrier (Em) itself, using our nebBVSE122k and nebDFT2k datasets. Remember, these barriers are for specific hops between two Li sites. So, for the GNNs, we got a bit creative and added a “dummy” atom at the midpoint of the hop to help the models focus on that specific pathway.

For the BVSE-NEB calculated barriers (the nebBVSE122k dataset), GNNs, particularly M3GNet, again showed better performance than classical models. We peeked into the Allegro GNN and found it was really paying attention to the area around that dummy midpoint atom. Smart, right?

Now, for the DFT-NEB data (nebDFT2k), things got interesting. Since DFT is more accurate but we had less data, we tried a transfer learning approach. We took GNNs pre-trained on the larger BVSE dataset and fine-tuned them on the DFT data. For classical models, we simply included the BVSE-calculated barrier as an extra feature. And guess what? This helping hand improved predictions! Surprisingly, a Random Forest model, beefed up with the BVSE barrier as a feature, actually gave the best accuracy for predicting the DFT-NEB barrier (MAE of 0.25 eV). It seems when data is scarce, well-engineered features for classical models can still pack a punch, even against pre-trained GNNs.

Round 3: The uMLIP Showdown – Can They Walk the Path?

This was the main event for me! We took our nebDFT2k and MPLiTrj datasets and unleashed four state-of-the-art uMLIPs: M3GNet, CHGNet, SevenNet, and MACE-MP-0. Our goal was to see if they could not only predict the energy barrier but also the actual path the Li-ion takes. One tricky bit: our DFT calculations used a slightly different setup (PBE functional) than the data these uMLIPs were originally trained on (which often includes PBE+U for transition metals, TMs). This matters because those +U corrections can change the predicted barrier, especially for materials with TMs.

So, we split our dataset into “with TM” and “without TM” systems. As expected, all uMLIPs did much better on the “without TM” stuff. MACE and SevenNet were the frontrunners here, with pretty impressive mean absolute errors (MAEs) of 0.07-0.08 eV for “without TM” systems – that’s getting really close to DFT accuracy! For “with TM” systems, errors were higher (MAE ~0.20-0.21 eV). M3GNet and CHGNet, unfortunately, tended to underestimate the barriers quite a bit. The old BVSE method, on the other hand, overestimated them, mostly because it assumes a rigid lattice.

Why the struggle with TM systems? Well, these uMLIPs, trained on PBE+U data, just weren’t great at predicting forces for our PBE-calculated TM systems. It makes sense – different underlying physics. This highlighted the need for fine-tuning!

Fine-Tuning for the Win!

We took MACE and SevenNet and fine-tuned them on a slice of our MPLiTrj data. The results were fantastic! Fine-tuning significantly slashed the errors for TM-containing systems. The fine-tuned SevenNet (SevenNet-ft) achieved an MAE of just 0.11 eV for these tricky systems! It also got much better at predicting forces and energies for both TM and non-TM materials. This shows that it’s not just about matching the computational scheme, but also about feeding the models more data on those high-energy transition states that ions pass through. This helps the uMLIPs get a better “feel” for the potential energy surface.

And what about predicting the actual migration path? The fine-tuned models also got much better here, predicting paths much closer to the DFT-calculated ones. This is a huge advantage over simpler structure-to-property models, as knowing the path can give us clues on how to tweak a material to make ions move even faster!

Putting it to Work: Screening for Better Battery Coatings

Okay, benchmarks are great, but can we use this in the real world? You bet! We took our fine-tuned SevenNet-ft model and went hunting for potential cathode coating materials for a Li10GeP2S12 (LGPS) solid-state battery with a LiCoO2 cathode. Coatings are vital to stop the electrolyte and cathode from having nasty reactions. We screened a bunch of Li-containing materials from the Materials Project database, looking for stable ones with good electrochemical properties and, crucially, low Li-ion percolation barriers calculated by SevenNet-ft.

And we found some promising candidates! Materials like Li3AlF6 and Li3PO4, which are known good ionic conductors, popped up with barrier predictions pretty close to reported values. We also identified some new potential heroes like LiYF4 and Li2ZnCl4. This really shows that these fine-tuned uMLIPs can be plugged into high-throughput screening pipelines and deliver reliable results, speeding up the discovery process immensely.

So, What’s the Big Picture?

Our journey through benchmarking these ML models has been pretty illuminating.

• GNNs for Percolation: They’re great for quickly sifting through large numbers of materials to find promising ionic conductors based on BVEL-calculated percolation barriers. For smaller datasets, direct BVEL might still be the way to go due to GNNs’ prediction errors, but for massive screenings, GNNs are a time-saver.
• Migration Barrier Prediction (Em): This is still a tough nut to crack with direct structure-to-property models, whether classical ML or GNNs. While they can help pre-screen, the accuracy isn’t quite at the “replace DFT entirely” level yet. Explicitly telling GNNs where to look (like our dummy atom trick) might be a way forward.
• uMLIPs are Stars: Fine-tuned uMLIPs, like SevenNet and MACE, are showing incredible promise. They can predict not just the barrier but the migration path with accuracy approaching DFT, especially after fine-tuning. This is a massive step up from older methods and way faster than full DFT-NEB. They can even be used to “precondition” a migration path for a final DFT polish, saving computational time.

Of course, there are always limitations. Our dataset, while comprehensive, is still biased towards certain types of migration (vacancy-based) and near-equilibrium structures. Future work could explore other mechanisms like interstitial or concerted (group) migration, which are super important in fast ionic conductors. Imagine training uMLIPs on data that includes these complex, correlated movements – that would be something else! Plus, just knowing the migration barrier isn’t the whole story for ionic conductivity; lattice dynamics and carrier concentration matter too. So, collecting more AIMD-derived conductivity data will be key.

But honestly, I’m incredibly excited. We’ve shown that these advanced ML tools, especially fine-tuned uMLIPs, are ready to become powerful allies in our quest for better energy storage materials. Our LiTraj dataset, I hope, will be a valuable springboard for even more innovation in this space, helping us all design materials with truly outstanding transport properties faster than ever before. The future of batteries is looking brighter, and a bit more intelligent, thanks to ML!

Source: Springer