src/streamlit_app.py

import os
import json
import time
from datetime import datetime
import numpy as np
import pandas as pd
import streamlit as st
import matplotlib.pyplot as plt
import seaborn as sns
import joblib

# ML imports
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor, ExtraTreesRegressor
from sklearn.preprocessing import StandardScaler, PolynomialFeatures
from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
from sklearn.metrics import mean_squared_error, r2_score

# SHAP
import shap

# -------------------------
# Config & paths
# -------------------------
st.set_page_config(page_title="AI Feature Universe Explorer — Advanced + SHAP", layout="wide")
DATA_DIR = os.getenv("DATA_DIR", "./data")
os.makedirs(DATA_DIR, exist_ok=True)
CSV_PATH = os.path.join(DATA_DIR, "flatfile_universe_advanced.csv")
META_PATH = os.path.join(DATA_DIR, "feature_metadata_advanced.json")
PDF_PATH = os.path.join(DATA_DIR, "annotated_bibliography.pdf")
ENSEMBLE_ARTIFACT = os.path.join(DATA_DIR, "ensemble_models.joblib")

# -------------------------
# Utility: generate advanced dataset if missing
# -------------------------
def generate_advanced_flatfile(
    n_rows=3000,
    random_seed=42,
    max_polynomial_new=60,
    global_variance_multiplier=1.0,
    variance_overrides=None,
):
    """
    Generates a large synthetic, physics-aligned dataset with many engineered features.
    Allows control of variability per feature (through variance_overrides) or globally
    (via global_variance_multiplier).

    Args:
        n_rows: number of samples
        random_seed: RNG seed
        max_polynomial_new: limit on number of polynomial expansion features
        global_variance_multiplier: multiplier applied to all default stddevs
        variance_overrides: dict mapping feature name or substring → stddev multiplier
    """
    np.random.seed(random_seed)
    os.makedirs(DATA_DIR, exist_ok=True)
    if variance_overrides is None:
        variance_overrides = {}

    # --- base natural features across 8 use cases (expanded)
    natural_feats = [
        "vibration_x","vibration_y","motor_current","rpm","bearing_temp","ambient_temp","lube_pressure","power_factor",
        "furnace_temp","tap_temp","slag_temp","offgas_co","offgas_co2","o2_probe_pct","c_feed_rate","arc_power","furnace_pressure","feed_time",
        "mold_temp","casting_speed","nozzle_pressure","cooling_water_temp","billet_length","chemical_C","chemical_Mn","chemical_Si","chemical_S",
        "roll_speed","motor_load","coolant_flow","exit_temp","strip_thickness","line_tension","roller_vibration",
        "lighting_intensity","surface_temp","image_entropy_proxy",
        "spectro_Fe","spectro_C","spectro_Mn","spectro_Si","time_since_last_sample",
        "batch_id_numeric","weight_input","weight_output","time_in_queue","conveyor_speed",
        "shell_temp","lining_thickness","water_flow","cooling_out_temp","heat_flux"
    ]
    natural_feats = list(dict.fromkeys(natural_feats))  # dedupe

    # helper: compute adjusted stddev
    def effective_sd(feature_name, base_sd):
        # exact name override
        if feature_name in variance_overrides:
            return float(variance_overrides[feature_name])
        # substring override
        for key, val in variance_overrides.items():
            if key in feature_name:
                return float(val)
        # fallback: scaled base
        return float(base_sd) * float(global_variance_multiplier)

    # helper sampling heuristics
    def sample_col(name, n):
        name_l = name.lower()
        if "furnace_temp" in name_l or name_l.endswith("_temp") or "tap_temp" in name_l:
            sd = effective_sd("furnace_temp", 50)
            return np.random.normal(1550, sd, n)
        if name_l in ("tap_temp","mold_temp","shell_temp","cooling_out_temp","exit_temp"):
            sd = effective_sd(name_l, 30)
            return np.random.normal(200 if "mold" not in name_l else 1500, sd, n)
        if "offgas_co2" in name_l:
            sd = effective_sd("offgas_co2", 4)
            return np.abs(np.random.normal(15, sd, n))
        if "offgas_co" in name_l:
            sd = effective_sd("offgas_co", 5)
            return np.abs(np.random.normal(20, sd, n))
        if "o2" in name_l:
            sd = effective_sd("o2_probe_pct", 1)
            return np.clip(np.random.normal(5, sd, n), 0.01, 60)
        if "arc_power" in name_l or "motor_load" in name_l:
            sd = effective_sd("arc_power", 120)
            return np.abs(np.random.normal(600, sd, n))
        if "rpm" in name_l:
            sd = effective_sd("rpm", 30)
            return np.abs(np.random.normal(120, sd, n))
        if "vibration" in name_l:
            sd = effective_sd("vibration", 0.15)
            return np.abs(np.random.normal(0.4, sd, n))
        if "bearing_temp" in name_l:
            sd = effective_sd("bearing_temp", 5)
            return np.random.normal(65, sd, n)
        if "chemical" in name_l or "spectro" in name_l:
            sd = effective_sd("chemical", 0.15)
            return np.random.normal(0.7, sd, n)
        if "weight" in name_l:
            sd = effective_sd("weight", 100)
            return np.random.normal(1000, sd, n)
        if "conveyor_speed" in name_l or "casting_speed" in name_l:
            sd = effective_sd("casting_speed", 0.6)
            return np.random.normal(2.5, sd, n)
        if "power_factor" in name_l:
            sd = effective_sd("power_factor", 0.03)
            return np.clip(np.random.normal(0.92, sd, n), 0.6, 1.0)
        if "image_entropy_proxy" in name_l:
            sd = effective_sd("image_entropy_proxy", 0.25)
            return np.abs(np.random.normal(0.5, sd, n))
        if "batch_id" in name_l:
            return np.random.randint(1000,9999,n)
        if "time_since" in name_l or "time_in_queue" in name_l:
            sd = effective_sd("time_since", 20)
            return np.abs(np.random.normal(30, sd, n))
        if "heat_flux" in name_l:
            sd = effective_sd("heat_flux", 300)
            return np.abs(np.random.normal(1000, sd, n))
        return np.random.normal(0, effective_sd(name_l, 1), n)

    # build DataFrame
    df = pd.DataFrame({c: sample_col(c, n_rows) for c in natural_feats})

    # timestamps & metadata
    start = pd.Timestamp("2025-01-01T00:00:00")
    df["timestamp"] = pd.date_range(start, periods=n_rows, freq="T")
    df["cycle_minute"] = np.mod(np.arange(n_rows), 80)
    df["meta_plant_name"] = np.random.choice(["Rourkela","Jamshedpur","VSP","Bokaro","Kalinganagar","Salem"], n_rows)
    df["meta_country"] = "India"

    # --- synthetic features: physics informed proxies
    df["carbon_proxy"] = df["offgas_co"] / (df["offgas_co2"] + 1.0)
    df["oxygen_utilization"] = df["offgas_co2"] / (df["offgas_co"] + 1.0)
    df["power_density"] = df["arc_power"] / (df["weight_input"] + 1.0)
    df["energy_efficiency"] = df["furnace_temp"] / (df["arc_power"] + 1.0)
    df["slag_foaming_index"] = (df["slag_temp"] * df["offgas_co"]) / (df["o2_probe_pct"] + 1.0)
    df["yield_ratio"] = df["weight_output"] / (df["weight_input"] + 1e-9)

    # rolling stats, lags, rocs for a prioritized set
    rolling_cols = ["arc_power","furnace_temp","offgas_co","offgas_co2","motor_current","vibration_x","weight_input"]
    for rc in rolling_cols:
        if rc in df.columns:
            df[f"{rc}_roll_mean_3"] = df[rc].rolling(3, min_periods=1).mean()
            df[f"{rc}_roll_std_5"] = df[rc].rolling(5, min_periods=1).std().fillna(0)
            df[f"{rc}_lag1"] = df[rc].shift(1).fillna(method="bfill")
            df[f"{rc}_roc_1"] = df[rc].diff().fillna(0)

    # interaction & polynomial-lite
    df["arc_o2_interaction"] = df["arc_power"] * df["o2_probe_pct"]
    df["carbon_power_ratio"] = df["carbon_proxy"] / (df["arc_power"] + 1e-6)
    df["temp_power_sqrt"] = df["furnace_temp"] * np.sqrt(np.abs(df["arc_power"]) + 1e-6)

    # polynomial features limited to first 12 numeric columns
    numeric = df.select_dtypes(include=[np.number]).fillna(0)
    poly_source_cols = numeric.columns[:12].tolist()
    poly = PolynomialFeatures(degree=2, interaction_only=False, include_bias=False)
    poly_mat = poly.fit_transform(numeric[poly_source_cols])
    poly_names = poly.get_feature_names_out(poly_source_cols)
    poly_df = pd.DataFrame(poly_mat, columns=[f"poly__{n}" for n in poly_names], index=df.index)
    keep_poly = [c for c in poly_df.columns if c.replace("poly__","") not in poly_source_cols]
    poly_df = poly_df[keep_poly].iloc[:, :max_polynomial_new] if len(keep_poly) > 0 else poly_df.iloc[:, :0]
    df = pd.concat([df, poly_df], axis=1)

    # PCA embeddings across numeric sensors
    scaler = StandardScaler()
    scaled = scaler.fit_transform(numeric)
    pca = PCA(n_components=6, random_state=42)
    pca_cols = pca.fit_transform(scaled)
    for i in range(pca_cols.shape[1]):
        df[f"pca_{i+1}"] = pca_cols[:, i]

    # KMeans cluster label for operating mode
    kmeans = KMeans(n_clusters=6, random_state=42, n_init=10)
    df["operating_mode"] = kmeans.fit_predict(scaled)

    # surrogate models
    surrogate_df = df.copy()
    surrogate_df["furnace_temp_next"] = surrogate_df["furnace_temp"].shift(-1).fillna(method="ffill")
    features_for_surrogate = [c for c in ["furnace_temp","arc_power","o2_probe_pct","offgas_co","offgas_co2"] if c in df.columns]
    if len(features_for_surrogate) >= 2:
        X = surrogate_df[features_for_surrogate].fillna(0)
        y = surrogate_df["furnace_temp_next"]
        rf = RandomForestRegressor(n_estimators=50, random_state=42, n_jobs=-1)
        rf.fit(X, y)
        df["pred_temp_30s"] = rf.predict(X)
    else:
        df["pred_temp_30s"] = df["furnace_temp"]

    if all(c in df.columns for c in ["offgas_co","offgas_co2","o2_probe_pct"]):
        X2 = df[["offgas_co","offgas_co2","o2_probe_pct"]].fillna(0)
        rf2 = RandomForestRegressor(n_estimators=50, random_state=1, n_jobs=-1)
        rf2.fit(X2, df["carbon_proxy"])
        df["pred_carbon_5min"] = rf2.predict(X2)
    else:
        df["pred_carbon_5min"] = df["carbon_proxy"]

    # safety indices & flags
    df["refractory_limit_flag"] = (df["lining_thickness"] < 140).astype(int)
    df["max_allowed_power_delta"] = np.clip(df["arc_power"].diff().abs().fillna(0), 0, 2000)

    # rule-based target
    df["ARC_ON"] = ((df["arc_power"] > df["arc_power"].median()) & (df["carbon_proxy"] < 1.0)).astype(int)
    df["prediction_confidence"] = np.clip(np.random.beta(2,5, n_rows), 0.05, 0.99)

    # clean NaN and infinite
    df.replace([np.inf, -np.inf], np.nan, inplace=True)
    df.fillna(method="bfill", inplace=True)
    df.fillna(0, inplace=True)

    # save CSV & metadata
    df.to_csv(CSV_PATH, index=False)
    meta = []
    for col in df.columns:
        if col in natural_feats:
            source = "natural"
        elif col.startswith("poly__") or col.startswith("pca_") or col in ["operating_mode"]:
            source = "advanced_synthetic"
        else:
            source = "synthetic"
        meta.append({
            "feature_name": col,
            "source_type": source,
            "linked_use_cases": ["All" if source!="natural" else "Mapped"],
            "units": "-",
            "formula": "see generator logic",
            "remarks": "auto-generated or simulated"
        })
    with open(META_PATH, "w") as f:
        json.dump(meta, f, indent=2)

    # annotated bibliography
    try:
        from fpdf import FPDF
        pdf = FPDF('P','mm','A4')
        pdf.add_page()
        pdf.set_font("Helvetica","B",14)
        pdf.cell(0,8,"Annotated Bibliography - Metallurgical AI (Selected Papers)", ln=True)
        pdf.ln(2)
        pdf.set_font("Helvetica","",10)
        pdf.cell(0,6,"Generated: " + datetime.utcnow().strftime("%Y-%m-%d %H:%M UTC"), ln=True)
        pdf.ln(4)
        bib_items = [
            ("A Survey of Data-Driven Soft Sensing in Ironmaking Systems","Yan et al. (2024)","Review of soft-sensors; supports gas proxies, lags, PCA."),
            ("Optimisation of Oxygen Blowing Process using RL","Ojeda Roldan et al. (2022)","RL for oxygen control; motivates surrogate predicted states & safety indices."),
            ("Analyzing the Energy Efficiency of Electric Arc Furnace","Zhuo et al. (2024)","Energy KPIs (kWh/t) motivate power_density & energy_efficiency features."),
            ("BOF/Endpoint prediction techniques","Springer (2024)","Endpoint prediction; supports temporal lags and cycle encoding."),
            ("Dynamic EAF modeling & slag foaming","MacRosty et al.","Physics priors for slag_foaming_index and refractory health modeling.")
        ]
        for title, auth, note in bib_items:
            pdf.set_font("Helvetica","B",11)
            pdf.multi_cell(0,6, f"{title} — {auth}")
            pdf.set_font("Helvetica","",10)
            pdf.multi_cell(0,5, f"Notes: {note}")
            pdf.ln(2)
        pdf.output(PDF_PATH)
    except Exception as e:
        with open(PDF_PATH.replace(".pdf",".txt"), "w") as tf:
            tf.write("Annotated bibliography generated. Install fpdf for PDF output.\n")

    return CSV_PATH, META_PATH, PDF_PATH

# -------------------------
# Ensure dataset exists
# -------------------------
if not os.path.exists(CSV_PATH) or not os.path.exists(META_PATH):
    with st.spinner("Generating synthetic features (this may take ~20-60s)..."):
        CSV_PATH, META_PATH, PDF_PATH = generate_advanced_flatfile(n_rows=3000, random_seed=42, max_polynomial_new=80)
        st.success(f"Generated dataset and metadata: {CSV_PATH}")

# -------------------------
# Load data & metadata (cached)
# -------------------------
@st.cache_data
def load_data(csv_path=CSV_PATH, meta_path=META_PATH):
    df_local = pd.read_csv(csv_path)
    with open(meta_path, "r") as f:
        meta_local = json.load(f)
    return df_local, pd.DataFrame(meta_local)

df, meta_df = load_data()

# -------------------------
# Sidebar filters & UI
# -------------------------
st.sidebar.title("Feature Explorer - Advanced + SHAP")
feat_types = sorted(meta_df["source_type"].unique().tolist())
selected_types = st.sidebar.multiselect("Feature type", feat_types, default=feat_types)
numeric_cols = df.select_dtypes(include=[np.number]).columns.tolist()

# -------------------------
# Main tabs
# -------------------------
st.title("Steel Authority of India Limited (SHAP-enabled)")
tabs = st.tabs([
    "Features",
    "Visualize",
    "Correlations",
    "Stats",
    "Ensemble + SHAP",
    "Target & Business Impact",
    "Bibliography"
])

# ----- Features tab
with tabs[0]:
    st.subheader("Feature metadata")
    filtered_meta = meta_df[meta_df["source_type"].isin(selected_types)]
    st.dataframe(filtered_meta[["feature_name","source_type","formula","remarks"]].rename(columns={"feature_name":"Feature"}), height=400)
    st.markdown(f"Total features loaded: **{df.shape[1]}**  |  Rows: **{df.shape[0]}**")

# ----- Visualize tab
with tabs[1]:
    st.subheader("Feature visualization")
    col = st.selectbox("Choose numeric feature", numeric_cols, index=0)
    bins = st.slider("Histogram bins", 10, 200, 50)
    fig, ax = plt.subplots(figsize=(8,4))
    sns.histplot(df[col], bins=bins, kde=True, ax=ax)
    ax.set_title(col)
    st.pyplot(fig)
    st.write(df[col].describe().to_frame().T)

# ----- Correlations tab
with tabs[2]:
    st.subheader("Correlation explorer")
    default_corr = numeric_cols[:20] if len(numeric_cols) >= 20 else numeric_cols
    corr_sel = st.multiselect("Select features (min 2)", numeric_cols, default=default_corr)
    if len(corr_sel) >= 2:
        corr = df[corr_sel].corr()
        fig, ax = plt.subplots(figsize=(10,8))
        sns.heatmap(corr, cmap="coolwarm", center=0, ax=ax)
        st.pyplot(fig)
    else:
        st.info("Choose at least 2 numeric features to compute correlation.")

# ----- Stats tab
with tabs[3]:
    st.subheader("Summary statistics (numeric features)")
    st.dataframe(df.describe().T.style.format("{:.3f}"), height=500)


# ----- Ensemble + SHAP tab
with tabs[4]:
    st.subheader("Autonomous Ensemble Modeling + SHAP Explainability")

    # --- Step 1: Basic UI selections ---
    target = st.selectbox("Target variable", numeric_cols, index=numeric_cols.index("furnace_temp") if "furnace_temp" in numeric_cols else 0)
    default_features = [c for c in numeric_cols if c != target][:60]
    features = st.multiselect("Model input features", numeric_cols, default=default_features)
    sample_size = st.slider("Sample rows for training", 500, min(4000, df.shape[0]), 1000, step=100)
    sub_df = df[features + [target]].sample(n=sample_size, random_state=42)
    X = sub_df[features].fillna(0)
    y = sub_df[target].fillna(0)

    # --- Step 2: Business / Process Objective selection ---
    st.markdown("### 🎯 Select Operational Objective")
    objective = st.selectbox(
        "Optimization Objective",
        [
            "Maximize Accuracy (R²)",
            "Minimize RMSE (Stable Control)",
            "Maximize Yield Ratio (EAF/Inventory)",
            "Minimize Energy Consumption (Efficiency)",
            "Balanced (Accuracy + Efficiency)"
        ],
        index=0
    )

    # --- Step 3: Auto-tuning with Optuna ---
    import optuna
    from sklearn.model_selection import cross_val_score

    st.markdown("### ⚙️ Auto Tuning in Progress")

    def objective_fn(trial):
        model_name = trial.suggest_categorical("model", ["RandomForest", "GradientBoosting", "ExtraTrees"])
        n_estimators = trial.suggest_int("n_estimators", 100, 600)
        max_depth = trial.suggest_int("max_depth", 3, 20)
        learning_rate = trial.suggest_float("learning_rate", 0.01, 0.3, log=True)

        if model_name == "RandomForest":
            model = RandomForestRegressor(n_estimators=n_estimators, max_depth=max_depth, n_jobs=-1)
        elif model_name == "GradientBoosting":
            model = GradientBoostingRegressor(n_estimators=n_estimators, learning_rate=learning_rate, max_depth=max_depth)
        else:
            model = ExtraTreesRegressor(n_estimators=n_estimators, max_depth=max_depth, n_jobs=-1)

        # Metric selection
        scoring_metric = "r2"
        if "RMSE" in objective:
            scoring_metric = "neg_root_mean_squared_error"

        score = cross_val_score(model, X, y, cv=3, scoring=scoring_metric).mean()
        return score

    if st.button("Run Auto Ensemble Optimization"):
        with st.spinner("Optimizing models... please wait (~20–60s)"):
            study = optuna.create_study(direction="maximize")
            study.optimize(objective_fn, n_trials=20)

            best_params = study.best_params
            st.success("✅ Best Auto-Tuned Model Found")
            st.json(best_params)

            # Build best model
            model_name = best_params.pop("model")
            if model_name == "RandomForest":
                model = RandomForestRegressor(**best_params)
            elif model_name == "GradientBoosting":
                model = GradientBoostingRegressor(**best_params)
            else:
                model = ExtraTreesRegressor(**best_params)
            model.fit(X, y)

            # Save model
            joblib.dump(model, ENSEMBLE_ARTIFACT)
            st.caption(f"Model saved: {ENSEMBLE_ARTIFACT}")

            # --- Auto Visualizations ---
            st.markdown("### 📈 Optimization History")
            fig_hist = optuna.visualization.matplotlib.plot_optimization_history(study)
            st.pyplot(fig_hist)

            # Predictions
            X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
            y_pred = model.predict(X_test)

            r2 = r2_score(y_test, y_pred)
            rmse = mean_squared_error(y_test, y_pred, squared=False)

            st.metric("R² Score", f"{r2:.3f}")
            st.metric("RMSE", f"{rmse:.3f}")

            # Scatter plot
            fig, ax = plt.subplots(figsize=(7,4))
            ax.scatter(y_test, y_pred, alpha=0.6)
            ax.plot([y_test.min(), y_test.max()], [y_test.min(), y_test.max()], "r--")
            ax.set_xlabel("Actual"); ax.set_ylabel("Predicted")
            st.pyplot(fig)

            # --- SHAP Explainability for Best Model ---
            st.markdown("### 🔍 SHAP Explainability (Auto Model)")
            explainer = shap.TreeExplainer(model)
            shap_values = explainer.shap_values(X_test.sample(300))
            fig_shap = plt.figure(figsize=(8,6))
            shap.summary_plot(shap_values, X_test.sample(300), show=False)
            st.pyplot(fig_shap)

            st.info("Auto tuning complete. Model performance and SHAP summary shown above.")

    
# -----  Target & Business Impact tab
with tabs[5]:
    st.subheader("Recommended Target Variables by Use Case")
    st.markdown("Each use case maps to a practical target variable that drives measurable business impact.")

    target_table = pd.DataFrame([
        ["Predictive Maintenance (Mills, Motors, Compressors)", "bearing_temp / time_to_failure", "Rises before mechanical failure; early warning", "₹10–30 L per asset/year"],
        ["Blast Furnace / EAF Data Intelligence", "furnace_temp / tap_temp", "Central control variable, linked to energy and quality", "₹20–60 L/year"],
        ["Casting Quality Optimization", "defect_probability / solidification_rate", "Determines billet quality; control nozzle & cooling", "₹50 L/year yield gain"],
        ["Rolling Mill Energy Optimization", "energy_per_ton / exit_temp", "Directly tied to energy efficiency", "₹5–10 L/year per kWh/t"],
        ["Surface Defect Detection (Vision AI)", "defect_probability", "Quality metric from CNN", "1–2 % yield gain"],
        ["Material Composition & Alloy Mix AI", "deviation_from_target_grade", "Predict deviation, suggest corrections", "₹20 L/year raw material savings"],
        ["Inventory & Yield Optimization", "yield_ratio (output/input)", "Linked to WIP and process yield", "₹1 Cr+/year"],
        ["Refractory & Cooling Loss Prediction", "lining_thickness / heat_loss_rate", "Predict wear for planned maintenance", "₹40 L/year downtime savings"]], columns=["Use Case", "Target Variable", "Why It’s Ideal", "Business Leverage"])

    st.dataframe(target_table,  use_container_width=True)

    st.markdown("---")
    st.subheader("Business Framing for Clients")
    st.markdown("These metrics show approximate annual benefits from small process improvements.")

    business_table = pd.DataFrame([
        ["Energy consumption", "400 kWh/ton", "₹35–60 L"],
        ["Electrode wear", "1.8 kg/ton", "₹10 L"],
        ["Refractory wear", "3 mm/heat", "₹15 L"],
        ["Oxygen usage", "40 Nm³/ton", "₹20 L"],
        ["Yield loss", "2 %", "₹50 L – ₹1 Cr"],
    ], columns=["Metric", "Typical Value (EAF India)", "5 % Improvement → Annual ₹ Value"])

    st.dataframe(business_table, use_container_width=True)
    st.info("These numbers are indicative averages; actual benefits depend on plant capacity and process efficiency.")

# -----  Bibliography tab 
with tabs[6]:
    st.subheader("Annotated Bibliography & Feature Justification")
    st.markdown("""
This section summarizes published research supporting the feature design and modeling choices.
    """)

    bib_data = [
        ("A Survey of Data-Driven Soft Sensing in Ironmaking Systems", "Yan et al. (2024)", "Supports gas proxies, lags, PCA for off-gas and temperature correlation."),
        ("Optimisation of Oxygen Blowing Process using RL", "Ojeda Roldan et al. (2022)", "Reinforcement learning for oxygen control; motivates surrogate predicted states & safety indices."),
        ("Analyzing the Energy Efficiency of Electric Arc Furnace", "Zhuo et al. (2024)", "Energy KPIs (kWh/t) motivate power_density & energy_efficiency features."),
        ("BOF/Endpoint Prediction Techniques", "Springer (2024)", "Endpoint prediction; supports temporal lags and cycle encoding."),
        ("Dynamic EAF Modeling & Slag Foaming", "MacRosty et al.", "Physics priors for slag_foaming_index and refractory health modeling."),
    ]

    bib_df = pd.DataFrame(bib_data, columns=["Paper Title", "Authors / Year", "Relevance to Feature Engineering"])
    st.dataframe(bib_df, use_container_width=True)

    st.markdown("""
**Feature-to-Research Mapping Summary:**
- Gas probes & soft-sensing → `carbon_proxy`, `oxygen_utilization`
- Power & energy proxies → `power_density`, `energy_efficiency`
- Temporal features → rolling means, lags, cycle progress indicators
- Surrogate features → `pred_temp_30s`, `pred_carbon_5min`
- PCA / clustering → operating mode compression
""")
# -------------------------
# Footer / Notes
# -------------------------
st.markdown("---")
st.markdown("**Notes:** This dataset is synthetic and for demo/prototyping. Real plant integration requires NDA, data on-boarding, sensor mapping, and plant safety checks before any control actions.")