Private. One row per simulated single laser track in IN718 (EOS M290 class): settings in, melt-pool size, energy terms, cooling and health out. Simulated with laserbeamFoamGPU (a CUDA port of LaserbeamFoam) on one RTX 5090. This upload is updated every 10 finished rows while the 100-row round runs; it currently holds 39 finished rows of 100 planned. The table has one row per finished run; the run IDs still to come are in plan/plan.parquet.

Use: pd.read_parquet("data/closure_meltpool_F1.parquet").query("status == 'success'"). QC failures (qc_fail) keep their outputs with qc_reasons; failures (fail) keep fail_reason.

Files

Path	What
`data/closure_meltpool_F1.parquet`	the dataset table (schema in `schema.json` and below)
`plan/plan.parquet`	the frozen plan: run IDs, knobs, split labels, F0 textbook baseline, decided before any run
`fields/<run_id>.npz`	per-row picture file: transverse cross-section (T, metal fraction, melted) at the steady plane and of the solidified track, centreline slice, top-surface T, 20,000 pool-boundary points (µm), grid axes, metadata JSON
`pictures/<run_id>.png`	quick-look image of each picture file
`cases/<run_id>.tar.zst`	full OpenFOAM case (all saved frames) of the anchors, ~1 % random rows and all failures (16 so far); `tar --zstd -xf`
`renders/`	cinematic renders (not part of the table)
`coverage_report.md`	counts per split, shape, regime and status; failure reasons
`checksums.sha256`	checksum of every dataset file
`docs/`	the brief (`scope.md`) and how the dataset is made (`DATASET_RUNBOOK.md`)

Renders:

renders/render_preview_F1-r00-s0-00002_cinematic.mp4

IN718 laser powder-bed fusion single tracks (EOS M290 class), one row per simulated melt pool, for training a fast melt-pool model. Brief: docs/scope.md; how it is made: docs/DATASET_RUNBOOK.md.

Rows: 39 of 100 planned (partial snapshot), schema F1-lbf-1.1, plan hash ba968102a46520d9, seed 20261020.
Simulator: laserbeamFoamGPU (CUDA port of LaserbeamFoam: two-phase VOF metal + argon, ray-traced Fresnel absorption with multiple reflections, Marangoni, recoil, HKL evaporation, latent heat), locked tag(s) dataset-lock-1; build(s): laserbeamFoamGPU 0.6.0-dev (git dataset-lock-1, CUDA 13.3.73, sm_120, Release); physicsFixes: boundaryKappaCp=on,scaledBarycentric=on.
Mesh: 4 um cells where the pool lives (3 um on the fine retry rung), graded outside.
Splits are fixed in the plan (split): anchor, train, test, ood (overhang angle 15-25 deg, never in train/test), active. Never re-split.
Failures are rows (status = fail, with fail_reason); QC failures keep their outputs (status = qc_fail, qc_reasons). Filter on status == "success" for clean training data.

How the inputs map to the solver

lambda_A (absorption multiplier): the solver has no absorptivity input (absorption comes from Fresnel optics per ray bounce). lambda_A scales the beam power fed to the ray tracer, P_effective_W = lambda_A * P_W. For a given pool shape this scales the absorbed power by lambda_A exactly (ray tracing is linear in power); it does not change the per-bounce Fresnel coefficient, so the reflected light scales too. Evaporation, recoil and pool shape respond to the extra heat.
f_M (stirring factor): the solver resolves the flow, so f_M multiplies the Marangoni coefficient dsigma/dT (plan section 1, user decision 2026-10-18).
Powder: resolved particles (seeded packing, 15-45 um log-normal); phi_achieved is what the packer reached in the 40 um layer (dense targets above ~0.50 are not reachable in a 40 um layer).
Neighbouring track: consolidated bead + fusion zone in the initial metal field, its heat as an instantaneous line source (F0 model) capped below T_s - 30 K, added to T_macro.
Delivered beam power: the solver seeds rays only within 1.5 r_b, so it delivers 98.9 % of the power it is given (P_delivered_W; all CPU and GPU runs, dev log F36). absorptivity is Q_abs_W / P_W (nominal).
Early stop: a run stops once the W and D means of its last three 50 um cross-section blocks agree within steady_tolerance (at least one cell). Rows started before 2026-10-21 03:05 IST used 2 %, later rows 6 % (steady_rule, steady_tolerance); the 2 % rows almost never stopped early. Most tracks are sized for exactly three blocks, so the rule only trims the end of a run.
Air cooling (h = 10 W/m^2 K): not in the solver; Q_conv_W is computed analytically on the metal surface and is far below 1 % of the absorbed power.

Known limits

W, D: melted envelope on solidified cross-sections in the steady window (NIST definition); W is the widest extent including the consolidated layer. L, volume, T_peak: instantaneous liquid pool.
Uncalibrated physics. Rows use the untuned material and model settings of configs/runs/campaign_base.yaml; against NIST AM-Bench 2022 the untuned model gives W +15 to +34 %, D -5 to -10 % (cases 0, 2.1, 2.2, 3.1; dev log F37). Calibration was postponed by the user. Every row carries solver_tag and config_fingerprint, so rows of a calibrated round can be told apart.
T_peak above T_b + 300 K is normal for evaporating keyholes; flag_T_peak_brief records the brief's rule and does not fail rows (dev log F9).
Pools larger than the affordable box (hot plate + hot neighbour + high Ke) are clipped by the cell budget; they fail the containment QC or are not steady (steady = false).
Bead height is measured on the final metal surface; bead_height_estimate_um is the mass balance.

Columns

Column	Type	Unit	Group	Meaning
`run_id`	str	-	id	Unique run ID, fixed in the plan before any run: F1-r-s-
`campaign`	str	-	id	Campaign name (F1)
`round`	int	-	id	Design round (0 = fixed design; 1-5 = active-learning rounds)
`shard`	int	-	id	Batch (machine) the run was assigned to
`design_index`	int	-	id	Global index in the 5,000-row design (anchors 1-50, train 51-2500, test 2501-3000)
`batch`	str	-	id	Planning batch that issued the run ID (e.g. b000)
`split`	str	-	id	anchor
`anchor_name`	str	-	id	For anchors: the reference case (e.g. NIST AMB2022-03 case 0)
`schema_version`	str	-	provenance	Schema version of this row
`plan_seed`	int	-	provenance	Master seed of the design
`code_version`	str	-	provenance	git describe of the dataset repo that made the row (-dirty if uncommitted)
`config_fingerprint`	str	-	provenance	Hash of every solver/case setting of the attempt that made the row
`solver`	str	-	provenance	Solver executable (laserbeamFoamGPU)
`solver_tag`	str	-	provenance	Locked solver tag (dataset-lock-1)
`solver_build`	str	-	provenance	The solver's Build line (version, git commit, CUDA, arch)
`physics_fixes`	str	-	provenance	laserbeamFoamGPU physicsFixes state (boundaryKappaCp, scaledBarycentric)
`machine`	str	-	provenance	Host name
`gpu`	str	-	provenance	GPU model
`cpu`	str	-	provenance	CPU model
`started_utc`	str	-	provenance	Start of the attempt that made the row (UTC, ISO 8601)
`finished_utc`	str	-	provenance	End of extraction (UTC, ISO 8601)
`attempts`	int	-	provenance	Solver attempts made for this run (retry ladder + requeues)
`rung`	str	-	provenance	Retry-ladder rung that made the row: base
`shape`	str	-	input	bulk
`has_neighbour`	bool	-	input	A previous track was just laid next to this one
`v_mm_s`	float	mm/s	input	Scan speed (Sobol, log scale)
`Ke_target`	float	-	input	Keyhole number the plan asked for (with A_nom = 0.6, no lambda_A)
`P_W`	float	W	input	Nominal laser power, from Ke_target (rows outside 80-380 W were discarded)
`P_effective_W`	float	W	input	Power fed to the ray tracer: lambda_A * P_W (see dataset card)
`P_delivered_W`	float	W	input	Beam power the solver delivers: rays are seeded within 1.5 r_b only, so P_effective_W * (1 - exp(-2 * 1.5^2)) = 0.989 P_effective_W (dev log F36)
`r_b_um`	float	um	input	Beam radius, 1/e^2 (D4sigma / 2)
`T_macro_K`	float	K	input	Part (plate) temperature before the track
`T_macro_block`	str	-	input	main (293-1100 K)
`h_prev_um`	float	um	input	Lateral distance to the neighbouring track (NaN: lone track)
`t_ret_ms`	float	ms	input	Age of the neighbouring track (NaN: lone track)
`lambda_A`	float	-	input	Absorption multiplier (uncertain knob)
`f_M`	float	-	input	Stirring factor, applied as a Marangoni dsigma/dT multiplier (plan section 1)
`phi_target`	float	-	input	Powder packing fraction asked for
`phi_achieved`	float	-	input	Packing fraction the packer reached in the 40 um layer
`n_particles`	int	-	input	Powder particles in the case (layer + powder below the surface)
`powder_seed`	int	-	input	Seed of the powder packing
`overhang_angle_deg`	float	deg	input	Overhang: angle of the sloped solid edge from horizontal
`edge_distance_um`	float	um	input	Overhang: lateral distance from the track centre to the top of the edge
`skin_thickness_um`	float	um	input	Skin over powder: solid sheet thickness under the layer
`wall_width_um`	float	um	input	Thin wall: wall width
`wall_offset_um`	float	um	input	Thin wall: track centre minus wall centre (lateral)
`layer_um`	float	um	input	Powder layer thickness (fixed, 40 um)
`material`	str	-	input	Material (IN718, configs/materials/in718.yaml)
`h_conv_W_m2K`	float	W/(m^2 K)	input	Air cooling coefficient (fixed, 10); applied analytically, see Q_conv_W
`T_pre_K`	float	K	geometry	Temperature the beam sees on arrival (T_macro + neighbour field at the track)
`gradT_lat_K_m`	float	K/m	geometry	Lateral temperature gradient dT/dz at the track (initial field)
`solid_fraction_below`	float	-	geometry	Solid (non-powder) metal fraction in a 3 r_b x 150 um box below the track
`cond_down_near`	float	-	geometry	Conductance descriptor: solid fraction along the down ray, near band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_mid`	float	-	geometry	Conductance descriptor: solid fraction along the down ray, mid band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_far`	float	-	geometry	Conductance descriptor: solid fraction along the down ray, far band (near 0-50, mid 50-150, far 150-300 um)
`cond_left_near`	float	-	geometry	Conductance descriptor: solid fraction along the left ray, near band (near 0-50, mid 50-150, far 150-300 um)
`cond_left_mid`	float	-	geometry	Conductance descriptor: solid fraction along the left ray, mid band (near 0-50, mid 50-150, far 150-300 um)
`cond_left_far`	float	-	geometry	Conductance descriptor: solid fraction along the left ray, far band (near 0-50, mid 50-150, far 150-300 um)
`cond_right_near`	float	-	geometry	Conductance descriptor: solid fraction along the right ray, near band (near 0-50, mid 50-150, far 150-300 um)
`cond_right_mid`	float	-	geometry	Conductance descriptor: solid fraction along the right ray, mid band (near 0-50, mid 50-150, far 150-300 um)
`cond_right_far`	float	-	geometry	Conductance descriptor: solid fraction along the right ray, far band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_left_near`	float	-	geometry	Conductance descriptor: solid fraction along the down-left ray, near band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_left_mid`	float	-	geometry	Conductance descriptor: solid fraction along the down-left ray, mid band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_left_far`	float	-	geometry	Conductance descriptor: solid fraction along the down-left ray, far band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_right_near`	float	-	geometry	Conductance descriptor: solid fraction along the down-right ray, near band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_right_mid`	float	-	geometry	Conductance descriptor: solid fraction along the down-right ray, mid band (near 0-50, mid 50-150, far 150-300 um)
`cond_down_right_far`	float	-	geometry	Conductance descriptor: solid fraction along the down-right ray, far band (near 0-50, mid 50-150, far 150-300 um)
`f0_A`	float	-	baseline	Absorptivity used by F0: A_nom (0.6) * lambda_A
`f0_W_um`	float	um	baseline	Eagar-Tsai width at f0_A * P, k and cp averaged from T_pre to T_l
`f0_D_um`	float	um	baseline	Eagar-Tsai depth (conduction; under-predicts keyholes)
`f0_L_um`	float	um	baseline	Eagar-Tsai length
`f0_Ke`	float	-	baseline	Keyhole number A B* / (30 pi) with f0_A and T_pre (Ke = 1: King's threshold)
`f0_D_size_um`	float	um	baseline	Depth the box was sized for (max of Eagar-Tsai and a keyhole-depth estimate)
`cell_um`	float	um	mesh	Uniform cell size in the melt-pool zone
`n_cells`	int	-	mesh	Cells in the mesh
`domain_length_um`	float	um	mesh	Domain length along the scan
`domain_half_width_um`	float	um	mesh	Domain half-width
`domain_depth_um`	float	um	mesh	Domain depth below the substrate top
`fine_depth_um`	float	um	mesh	Depth of the uniform fine zone
`fine_half_width_um`	float	um	mesh	Half-width of the uniform fine zone
`track_length_um`	float	um	mesh	Planned track length (the run may stop earlier, steady window)
`W_um`	float	um	output	Melt-pool width: widest melted extent of the solidified track cross-sections
`W_std_um`	float	um	output	Std of W over the cross-sections of the steady window
`D_um`	float	um	output	Melt-pool depth below the substrate top (y = 0), solidified cross-sections
`D_std_um`	float	um	output	Std of D over the steady window
`L_um`	float	um	output	Liquid pool length (T >= T_l, largest connected region), mean over steady frames
`L_std_um`	float	um	output	Std of L over the steady frames
`W_top_um`	float	um	output	Melted width at the powder-layer top (y = 40 um); 0 if the bead is lower
`D_from_powder_top_um`	float	um	output	Depth measured from the powder-layer top: D + 40 um
`remelt_depth_um`	float	um	output	Depth of melting into metal that was solid before the track (not powder)
`volume_um3`	float	um^3	output	Liquid pool volume, mean over steady frames
`volume_std_um3`	float	um^3	output	Std of the liquid volume
`D_over_W`	float	-	output	D / W
`T_peak_K`	float	K	output	Peak metal temperature, mean over steady frames
`T_peak_std_K`	float	K	output	Std of the peak temperature over steady frames
`T_peak_max_K`	float	K	output	Highest metal temperature in any frame
`T_surface_p99_K`	float	K	output	99th percentile of hot surface-metal temperature, max over frames
`regime`	str	-	output	conduction (D/W < 0.5)
`bead_height_um`	float	um	output	Bead top above the substrate top, measured on the final metal surface
`bead_height_estimate_um`	float	um	output	Bead height by mass conservation: 1.5 phi layer (W + 2 r_b) / W
`bead_method`	str	-	output	How bead_height_um was obtained
`Q_abs_W`	float	W	energy	Laser power absorbed by the metal (ray tracer), mean over the steady window
`Q_abs_std_W`	float	W	energy	Std of the absorbed power per step
`absorptivity`	float	-	energy	Q_abs_W / P_W (includes lambda_A)
`Q_evap_W`	float	W	energy	Evaporation loss (solver's HKL term)
`Q_rad_W`	float	W	energy	Radiation loss (solver's term)
`Q_conv_W`	float	W	energy	Air cooling h = 10 W/(m^2 K) on the metal surface, computed analytically (not in the solver)
`Q_net_W`	float	W	energy	Q_abs - Q_evap - Q_rad - Q_conv
`Q_walls_W`	float	W	energy	Heat conducted in through the domain walls (negative: out)
`dE_stored_W`	float	W	energy	Rate of change of stored (sensible + latent) energy
`energy_balance_err_pct`	float	%	energy	(dE/dt - sources) / laser, steady window
`energy_balance_err_run_pct`	float	%	energy	Same over the whole run (the QC test)
`Ke_measured`	float	-	energy	Keyhole number with the solver's absorptivity instead of f0_A
`G_K_m`	float	K/m	cooling	Temperature gradient on the trailing liquidus surface (median)
`R_m_s`	float	m/s	cooling	Solidification speed v cos(theta) on the trailing liquidus (median)
`cooling_rate_K_s`	float	K/s	cooling	G R on the trailing liquidus (median)
`G_over_R_K_s_m2`	float	K s/m^2	cooling	G / R on the trailing liquidus (median)
`G_tail_surface_K_m`	float	K/m	cooling	G at the surface tail of the pool on the centreline
`R_tail_surface_m_s`	float	m/s	cooling	R at the surface tail of the pool on the centreline
`steady`	bool	-	window	W and D of three successive 50 um blocks agreed within 2 %
`early_stopped`	bool	-	window	The run was stopped as soon as it was steady
`steady_rule`	str	-	window	Early-stop rule the row ran under: last3_blocks = W and D block means of the last 3 blocks agree within steady_tolerance (at least one cell)
`steady_tolerance`	float	-	window	Relative tolerance of that rule (0.02 for rows started before 2026-10-21 03:05 IST, then 0.06)
`steady_blocks`	int	-	window	Cross-section blocks in the window at the end (NaN: not recorded, 1.0 rows)
`steady_x_start_um`	float	um	window	Start of the cross-section window (from the domain inlet)
`steady_x_end_um`	float	um	window	End of the cross-section window
`steady_n_planes`	int	-	window	Cross-sections averaged for W and D
`steady_t_start_s`	float	s	window	Start of the time window for L, volume, T_peak, energy
`steady_t_end_s`	float	s	window	End of that time window
`steady_n_frames`	int	-	window	Frames averaged for L, volume, T_peak
`spatter_loss_frac`	float	-	window	Metal lost through the open top (spatter) by the end, fraction of initial metal
`status`	str	-	health	success
`fail_reason`	str	-	health	Why the run failed (solver, extraction or retry ladder exhausted)
`qc_pass`	bool	-	health	All per-run QC checks passed
`qc_reasons`	str	-	health	Failed QC checks, '; '-separated
`qc_energy_ok`	bool	-	health
`qc_containment_ok`	bool	-	health	Pool never touched a wall and stayed in the fine zone
`qc_mesh_ok`	bool	-	health	W / cell >= 10
`qc_temperature_ok`	bool	-	health	No NaN, no undershoot below T0 - 50 K, no runaway above 8,000 K
`cells_across_W`	float	-	health	W / cell size
`flag_T_peak_brief`	bool	-	health	Brief's rule T_peak <= T_b + 300 K holds (reported only: every evaporating keyhole breaks it, dev log F9)
`flag_T_peak_hkl`	bool	-	health	T_peak <= T(HKL flux = beam peak intensity) + 300 K holds (proposed replacement, reported only)
`solver_steps`	int	-	health	Time steps taken
`physical_time_s`	float	s	health	Simulated time reached
`mean_dt_s`	float	s	health	Mean time step
`solver_wall_s`	float	s	health	Solver wall time of the final attempt
`row_wall_s`	float	s	health	Wall time of the final attempt: build + solve + extraction
`concurrent_jobs`	int	-	health	Rows running on the GPU when this one started (incl. itself)
`gpu_mem_peak_MiB`	float	MiB	health	Peak GPU memory of the solver process (nvidia-smi)
`host_rss_peak_GB`	float	GB	health	Peak host memory of the solver process
`fields_file`	str	-	files	Picture file, relative to the dataset folder (fields/.npz)
`fields_sha256`	str	-	files	sha256 of the picture file
`case_kept`	bool	-	files	The full OpenFOAM case was kept (anchors, ~1 % of rows, all failures)
`case_path`	str	-	files	Where the kept case is (relative to the dataset folder), else empty

Downloads last month: 4

Total file size:

28 GB