Neural Networks vs BDTs in Particle Physics: Which and When

July 26, 2026 · 5 min read

Choosing between a boosted decision tree and a neural network is one of the first real decisions you face when building a classifier for a new analysis. Both tools appear everywhere in HEP, both have strong track records, and the documentation for both is excellent — so it can feel arbitrary which one to reach for. It isn't arbitrary. The structure of your data makes the choice fairly systematic, and getting it right saves you weeks of tuning something that was never going to work well in the first place.

Start with Your Data, Not the Algorithm

The single most important question is: what does one event look like as input?

If your event is a fixed-length row of high-level reconstructed quantities — lepton pT, MET, jet multiplicity, invariant masses, angular separations — you have tabular data. If your event is a variable-length collection of objects (all the tracks, all the calorimeter cells, a jet's full constituent list), or an image, or a graph, you have structured or low-level data. That distinction almost entirely determines which tool will serve you better.

When BDTs Are the Right Call

For tabular ntuples of engineered observables, gradient-boosted decision trees — XGBoost, LightGBM, or TMVA's BDTG — remain extremely competitive with deep networks and are often easier to work with.

Why BDTs handle tabular data so well

• They learn well in low-data regimes. Physics analyses often have limited signal simulation. BDTs generalize reliably with training sets that would leave a deep network undertrained or prone to overtraining — the ML equivalent of overfitting a chi-square with too many free parameters.
• Feature engineering is rewarded. When you add a physically motivated combination — say, the collinear mass or a dedicated angular variable — a BDT picks it up immediately. Deep networks can in principle learn those combinations, but they need more data and more tuning to do so.
• Hyperparameter tuning is fast and interpretable. Depth, learning rate, number of trees: scanning these is cheap, and the effect of each knob is intuitive. Variable importance scores also map directly onto physics intuition, letting you cross-check that the classifier is using what you expect.
• Training and inference are fast. For analyses running over many systematic variations or doing blinded optimization studies, this matters.

If your input is a vector of thirty reconstructed quantities from your ntuple, start here. The neural network vs BDT debate often ends at this step.

When Deep Networks Are the Right Call

The landscape shifts once you move away from fixed-length, hand-crafted feature vectors.

Low-level and raw inputs

When you hand the classifier raw four-vectors for all jets, all tracks, or all calorimeter deposits, you are asking it to do physics — to discover the invariant masses, the angular correlations, the substructure — from scratch. BDTs cannot do this. They require a fixed input length and have no built-in notion of permutation invariance or variable-length sets. Architectures like particle flow networks, transformers operating on sets of particles, or graph neural networks on track-vertex graphs are built for exactly this structure. Deep learning in particle physics covers these architectures in detail.

Jet tagging and image-like inputs

Jet images — discretized calorimeter deposits projected onto an η–φ grid — are a natural home for convolutional networks. The translation and rotation symmetries of the grid match the inductive biases CNNs were designed for. A BDT on the raw pixel values would see an enormous input space with no awareness of spatial relationships. Deep networks are the clear choice here.

When labeled data is abundant

Deep networks tend to repay scale: more training examples, more expressive architectures, better performance. In contexts where you can generate large simulation samples — or use data-driven methods to augment your training set — a deep network can outperform a well-tuned BDT by exploiting subtle correlations the BDT cannot reach with fixed-depth trees.

A Practical Decision Checklist

Run through these questions before choosing:

1. Is my input a fixed-length vector of reconstructed quantities? → Start with a BDT.
2. Is my input variable-length, a set of particles, a graph, or an image? → Use an architecture designed for that structure.
3. Is my signal sample small or my MC statistics limited? → Favor BDTs; they generalize with less data.
4. Am I doing jet substructure or low-level tagging? → Deep networks, ideally with an appropriate inductive bias.
5. Do I need fast iteration, easy debugging, or interpretable feature importance? → BDTs give you this more naturally.

If you want to go deeper on both tools and understand how to tune them for real LHC analyses, the full HEP ML course walks through both methods with working code. You can also try the free Module 1 to see the approach before committing.

The One-Line Takeaway

Match your tool to your data structure: BDTs for engineered tabular features, deep networks for low-level, structured, or image-like inputs — and you will spend your time doing physics rather than fighting your classifier.

References

Baldi, P., Sadowski, P., & Whiteson, D. (2014). Searching for exotic particles in high-energy physics with deep learning. Nature Communications, 5, 4308. arXiv:1402.4735.

Guest, D., Cranmer, K., & Whiteson, D. (2018). Deep Learning and its Application to LHC Physics. Annual Review of Nuclear and Particle Science, 68, 161-181. arXiv:1806.11484.