Glossary

Plain-language definitions of every technical term used in this documentation. If you are new to CFD or HPC, read this section first.

kLa (mass-transfer coefficient) The rate at which oxygen dissolves from the air into the liquid in the bag, measured per unit time. Higher kLa = better mixing and faster oxygenation. kLa_25 is the value when the liquid has reached 25 % of full oxygen saturation and is the standard industrial metric. Units: h⁻¹ (converted from the solver's internal non-dimensional rate via 3600 / T_bio, so it is directly comparable to literature values such as Kim et al. 2024).

Rocking period (T) The time it takes the bag to complete one full back-and-forth rock. Related to rocking frequency: T = 2π / ω_b. Shorter period = faster rocking.

Angular frequency (ω_b, omega_b) How fast the bag rocks, measured in radians per second (rad/s). 1 Hz = 2π ≈ 6.28 rad/s. A typical bioreactor runs at 0.3–2 Hz.

Non-dimensional time Simulation time is scaled by the characteristic sloshing timescale T_bio = L / U_bio, where L is the bag length and U_bio is the typical sloshing velocity. This makes results independent of the exact bag size or fluid speed, so one non-dim time unit means roughly "one sloshing timescale has passed." Physical time in seconds = t × T_bio.

Fidelity / grid level How fine the computational mesh (grid) is. Grid size = 2^fidelity × 2^fidelity cells. Higher fidelity = more cells = more accurate results, but slower and more memory-intensive. Fidelity 3 (8×8) takes seconds; fidelity 7 (128×128) takes hours. See the Fidelity guide for a full table.

Checkpoint / checkpoint restart A snapshot of the full simulation state (fluid velocity, volume fraction, dissolved oxygen, etc.) saved to a binary file called checkpoint.dump. A later simulation can restore this snapshot and continue from that exact point, skipping the warm-up phase. This is used in chained sweeps to save 70–90 % of compute time.

Segment One SLURM job in a chained sweep. Each segment is a complete, self-contained simulation that starts either from scratch (segment 0) or from the checkpoint left by the previous segment.

Chain A sequence of segments linked by checkpoint restart. Each segment automatically submits its successor when it completes successfully (self-submitting chain). No external scheduler bookkeeping is required.

n_mix_cycles The number of rocking cycles to run before injecting oxygen. Used to let the flow reach a steady state before taking measurements. Typical value: 80 cycles for a fresh start, 10 for a checkpoint restart (flow is already developed).

t_buffer The duration (in non-dim time) of the kLa measurement window after oxygen injection. Larger t_buffer gives a longer average and more stable kLa estimate, at the cost of more compute time. Rule of thumb: t_buffer > ln(2) / kLa_nd_expected, where kLa_nd is the solver's internal non-dimensional rate (not the h⁻¹ value reported in results.json — t_buffer is itself a non-dim duration, so it must be sized against the non-dim rate). At fidelity 7, kLa_nd ≈ 0.03–0.1, so t_buffer = 30–50 is sufficient. At fidelity 5, kLa_nd ≈ 0.1–0.5, so t_buffer = 10–30 suffices.

Superellipse (geometry.n) The mathematical shape of the bag cross-section. n=2 is a standard ellipse; n=8 or more looks like a rounded rectangle. The shape is controlled by geometry.a (half-width), geometry.b (half-height), and geometry.n (roundness).

Surrogate model A fast mathematical approximation of kLa built by fitting to existing simulation data. Once trained, it can predict kLa for any parameter combination in milliseconds (vs. hours for a real simulation). Used in Bayesian optimisation to guide where to sample next.

Bayesian optimisation (BO) An algorithm that iteratively chooses which parameter combination to simulate next, based on a balance between exploring unknown regions (exploration) and refining around the best results found so far (exploitation). Uses the surrogate to evaluate candidates cheaply. The expected improvement (EI) acquisition function quantifies how promising each candidate is.

DoE (Design of Experiments) An initial set of simulation runs that covers the parameter space broadly before optimisation begins. Used to build the first surrogate model. Common strategy: Latin Hypercube Sampling (LHS), which spreads points evenly.

SLURM The job scheduler on OSCAR (Brown University HPC). You submit a job with sbatch; SLURM queues it and runs it on a compute node when resources are available. Jobs communicate via environment variables and output files — never interact with SLURM interactively from a login node.

HPC / OSCAR High-Performance Computing cluster at Brown University. Always verify you are on a compute node (not a login node) before running expensive simulations. Login nodes are shared; compute nodes are allocated exclusively per job.