{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "\n# Vector MFML for Dipole Moments\nNow we move on to an example of creating an MFML model for vector properties.\nThis example trains a 5-fidelity MFML model using\nVectorKRR to directly predict electronic\ndipole moment vectors\n(3 Cartesian components) for nitrophenol from the QeMFi dataset.\n\nWe will compares the vector MFML against 3 independent scalar MFML models (one per\ndipole component) to check if the vector MFML does add any improvement at all.\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Imports\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import numpy as np\nimport matplotlib.pyplot as plt\n\nfrom mfml_qc.datasets import load_nitrophenol_data\nfrom mfml_qc.representations import compute_flat_coulomb\nfrom mfml_qc.mfml import ModelMFML\nfrom mfml_qc.utils import top_down_subsetting\nfrom mfml_qc.vector_krr import CoupledVectorKRR\nfrom mfml_qc.vector_mfml import VectorMFML\nfrom mfml_qc.rpr import RPRVectorKRR" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Load and Split Data\nWe use the QeMFi nitrophenol dataset with 5 SCF fidelity levels.\nThe first 12,000 geometries form the training pool; the remaining\n3,000 are held out for testing against the highest-fidelity dipole\nvectors.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "dataset = load_nitrophenol_data()\nR = dataset[\"R\"]\nZ = dataset[\"Z\"]\nX_CM = dataset[\"X_CM\"]\nDPe = dataset[\"DPe\"] # (15000, 5, 3)\n\nN_TRAIN = 12000\nN_TEST = 3000\nNUM_FIDS = 5\n\nX_train_parent_cm = X_CM[:N_TRAIN]\nX_test_cm = X_CM[N_TRAIN : N_TRAIN + N_TEST]\nX_train_parent_flat = R[:N_TRAIN].reshape(N_TRAIN, -1)\nX_test_flat = R[N_TRAIN : N_TRAIN + N_TEST].reshape(N_TEST, -1)\nDPe_train = DPe[:N_TRAIN]\nDPe_test_ref = DPe[N_TRAIN : N_TRAIN + N_TEST, -1, :]" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Build Hierarchy for Vector Targets\nWe construct the multi-fidelity hierarchy manually since\n``build_hierarchy_arrays`` expects scalar targets. Each fidelity\nlevel's dipole vectors are mean-centred independently.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "y_trains = np.zeros(NUM_FIDS, dtype=object)\nindexes = np.zeros(NUM_FIDS, dtype=object)\nmeans = np.zeros((NUM_FIDS, 3))\n\nfor i in range(NUM_FIDS):\n vals = DPe_train[:, i, :]\n means[i] = np.mean(vals, axis=0)\n y_trains[i] = vals - means[i]\n indexes[i] = np.column_stack([np.arange(N_TRAIN), np.arange(N_TRAIN)])\n\nn_trains_target = [\n 4096,\n 2048,\n 1024,\n 512,\n 256,\n] # 256 at target fidelity 4096 at the lowest\nsubset_y_trains, subset_indexes = top_down_subsetting(\n y_trains, indexes, n_trains_target, seed=42\n)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Train Vector MFML\nA single ``VectorKRR`` instance is passed as the ``base_estimator``.\n``ModelMFML`` deep-copies it for each of the 2N\u22121 = 9 sub-models.\nPrediction uses the default MFML sub-model summation\nbecause ``ModelMFML.predict``\nworks on scalar outputs.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "base_model = CoupledVectorKRR(\n kernel_type=\"gaussian\",\n sigma=250.0,\n reg=1e-10,\n val_fraction=0.15,\n max_opt_iter=500,\n)\nbase_model = RPRVectorKRR(base_model, Z, compute_flat_coulomb)\n\nmfml_model = VectorMFML(base_estimator=base_model)\nmfml_model.train(\n X_train_parent=X_train_parent_flat,\n y_trains=subset_y_trains,\n indexes=subset_indexes,\n)\n\nDPe_pred = mfml_model.predict(X_test_flat) + means[-1]" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Scalar MFML Baseline\nFor comparison, we train 3 independent scalar MFML models \u2014 one per\nCartesian component of the dipole moment. With the isotropic kernel\nthis is mathematically equivalent to the vector formulation, but\nrequires managing 3 separate model objects.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "DPe_pred_scalar = np.zeros((N_TEST, 3))\n\nfor c in range(3):\n y_trains_c = np.zeros(NUM_FIDS, dtype=object)\n means_c = np.zeros(NUM_FIDS)\n for i in range(NUM_FIDS):\n vals = DPe_train[:, i, c]\n means_c[i] = np.mean(vals)\n y_trains_c[i] = vals - means_c[i]\n\n subset_y_c, subset_idx_c = top_down_subsetting(\n y_trains_c, indexes, n_trains_target, seed=42\n )\n\n mfml_c = ModelMFML(kernel=\"gaussian\", sigma=250.0, reg=1e-10, p_bar=False)\n mfml_c.train(\n X_train_parent=X_train_parent_cm,\n y_trains=subset_y_c,\n indexes=subset_idx_c,\n )\n preds_c = mfml_c.predict(X_test=X_test_cm, optimiser=\"default\") + means_c[-1]\n DPe_pred_scalar[:, c] = preds_c" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Evaluate\nWe compute component-wise MAE, overall MAE, and cosine similarity\nbetween predicted and reference dipole vectors.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "def evaluate(pred, true):\n comp_mae = np.mean(np.abs(pred - true), axis=0)\n overall = np.mean(np.abs(pred - true))\n mag_p, mag_t = np.linalg.norm(pred, axis=-1), np.linalg.norm(true, axis=-1)\n dot = np.sum(pred * true, axis=-1)\n cos_sim = np.clip(dot / np.maximum(mag_p * mag_t, 1e-12), -1.0, 1.0)\n return comp_mae, overall, cos_sim\n\n\ncomp_vec, mae_vec, cos_vec = evaluate(DPe_pred, DPe_test_ref)\ncomp_sca, mae_sca, cos_sca = evaluate(DPe_pred_scalar, DPe_test_ref)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Comparison Plot\nThree-panel figure: dipole magnitude parity (left), component-wise\nMAE bar chart (centre), and cosine similarity histograms (right).\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(16, 5))\n\nmag_true = np.linalg.norm(DPe_test_ref, axis=-1)\nmag_vec = np.linalg.norm(DPe_pred, axis=-1)\nmag_sca = np.linalg.norm(DPe_pred_scalar, axis=-1)\n\nax1.scatter(\n mag_true,\n mag_sca,\n s=2,\n alpha=0.35,\n color=\"#ff7f0e\",\n label=f\"Scalar MFML (MAE: {np.mean(np.abs(mag_sca - mag_true)):.4f})\",\n)\nax1.scatter(\n mag_true,\n mag_vec,\n s=2,\n alpha=0.35,\n color=\"#1f77b4\",\n label=f\"Vector MFML (MAE: {np.mean(np.abs(mag_vec - mag_true)):.4f})\",\n)\nlims = [\n min(mag_true.min(), mag_sca.min(), mag_vec.min()),\n max(mag_true.max(), mag_sca.max(), mag_vec.max()),\n]\nax1.plot(lims, lims, \"r--\", lw=1.5)\nax1.set_xlabel(\"True |\u03bc| (Debye)\")\nax1.set_ylabel(\"Predicted |\u03bc| (Debye)\")\nax1.set_title(\"Dipole Magnitude Parity\")\nax1.legend(markerscale=5, fontsize=8)\n\ncomp_labels = [\"\u03bc_x\", \"\u03bc_y\", \"\u03bc_z\", \"Overall\"]\nvec_maes = list(comp_vec) + [mae_vec]\nsca_maes = list(comp_sca) + [mae_sca]\nx_pos = np.arange(len(comp_labels))\nw = 0.35\nax2.bar(x_pos - w / 2, vec_maes, w, label=\"Vector MFML\", color=\"#1f77b4\")\nax2.bar(x_pos + w / 2, sca_maes, w, label=\"Scalar MFML\", color=\"#ff7f0e\")\nax2.set_xticks(x_pos)\nax2.set_xticklabels(comp_labels)\nax2.set_ylabel(\"MAE (Debye)\")\nax2.set_title(\"Component-wise MAE\")\nax2.legend(fontsize=8)\n\nax3.hist(\n cos_sca,\n bins=60,\n color=\"#ff7f0e\",\n edgecolor=\"none\",\n alpha=0.5,\n label=f\"Scalar (mean: {np.mean(cos_sca):.4f})\",\n)\nax3.hist(\n cos_vec,\n bins=60,\n color=\"#1f77b4\",\n edgecolor=\"none\",\n alpha=0.5,\n label=f\"Vector (mean: {np.mean(cos_vec):.4f})\",\n)\nax3.set_xlabel(\"Cosine Similarity\")\nax3.set_ylabel(\"Count\")\nax3.set_title(\"Directional Alignment\")\nax3.legend(fontsize=8)\n\nplt.tight_layout()\nplt.show()" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.10.22" } }, "nbformat": 4, "nbformat_minor": 0 }