awkward

Awkward Array

Overview

Awkward Array provides NumPy-like idioms for variable-length, nested, and record-structured data — the natural shape of HEP event data. Version 2.x (current) has a substantially different API from 1.x; do not mix patterns.

When to Use

• Reading TTrees via uproot (returns ak.Array automatically)
• Building physics-object records and applying per-object selection
• Computing pair quantities (invariant mass, deltaR) via combinatorics
• Broadcasting per-event scalars to per-object arrays for weighted fills
• Preparing arrays for histogram filling or NumPy interop

Key Concepts

Concept	Notes
ak.type(arr)	Inspect type — e.g., var * {pt: float64}
OptionType	None values from pad_none or firsts on empty events
Behaviors	vector.register_awkward() adds 4-vector methods to records
ak.num(arr)	Per-event count (number of jets per event, etc.)

Canonical Patterns

Build a record from columns:

import awkward as ak
jets = ak.zip({"pt": events["jet_pt"], "eta": events["jet_eta"],
               "phi": events["jet_phi"], "mass": events["jet_m"]})

Filter objects, then events:

good_jets = jets[jets.pt > 25_000]                    # per-object mask, MeV
events_ok = good_jets[ak.num(good_jets) >= 2]         # per-event mask

All unique pairs per event:

combos = ak.combinations(jets, 2, axis=1)
j1, j2 = ak.unzip(combos)

Leading object per event (keepdims is required for slicing):

lead_idx = ak.argmax(jets.pt, axis=1, keepdims=True)
lead = ak.firsts(jets[lead_idx])   # None for empty events

Add a derived field:

jets = ak.with_field(jets, jets.pt / 1000, "pt_gev")

Broadcast per-event weight to per-object for weighted fills:

flat_w = ak.flatten(ak.broadcast_arrays(events.weight, jets.pt)[0])
flat_pt = ak.flatten(jets.pt)
h.fill(pt=ak.to_numpy(flat_pt), weight=ak.to_numpy(flat_w))

Gotchas

• firsts vs [:, 0]: firsts returns None for empty events; [:, 0] raises. Use firsts whenever collections may be empty.
• flatten depth: Default axis=1 flattens one level. axis=None flattens everything — rarely correct.
• option propagation: Once OptionType enters, operations propagate None. Call ak.drop_none or ak.fill_none before NumPy interop.
• to_numpy on ragged: Fails unless regular. Always flatten or select a fixed depth first.
• Behavior registration is global: Call vector.register_awkward() once at module level; it mutates the global dict.
• 2.x breaking changes from 1.x: ak.ArrayBuilder → ak.from_iter; many behavior patterns changed; do not copy 1.x examples verbatim.

Interop

• uproot: tree.arrays() returns ak.Array — no conversion needed
• vector: register_awkward() gives records with pt/phi/eta/mass Lorentz methods
• hist: Flatten arrays before Hist.fill(); broadcast weights first
• coffea: NanoEvents columns are ak.Array; all coffea processors consume awkward natively

Docs

https://awkward-array.org/doc/main/