fix: correct energy calibration from Am-241 peaks

- Am-241 peaks at ch 134=13.8, ch 163=17.8, ch 218=26.3 keV
- Calibration: E = 0.397 + 0.0995 × ch (previously wrong)
- Kurie fit endpoint 425 keV vs Q=297.5 keV — spectrum extends
  beyond MCA range (407 keV), simple extrapolation unreliable
- Proper endpoint extraction requires theory model + detector
  convolution (fitter module), not simple Kurie plot

exp_data/raw_reader.py

@@ -142,8 +142,8 @@ def calibrate_from_am241(
 ) -> tuple[float, float]:
     """Calibrate energy scale using Am-241 calibration source.
 
-    Reads the Am-241 spectrum, finds prominent peaks, matches them
+    Matches known Am-241 gamma lines to detected peaks in the
-    to known gamma lines, and fits a linear energy calibration.
+    low-energy region (ch < 300) and fits a linear calibration.
 
     Parameters
     ----------
@@ -151,7 +151,7 @@ def calibrate_from_am241(
         Path to the Am-241 binary file (CAM1.DAT or CAM2.DAT).
     known_peaks
         Mapping of energy_keV -> expected channel. If None, auto-detects
-        Am-241 peaks at 13.8, 17.8, 26.3, 59.54 keV.
+        Am-241 peaks at 13.8, 17.8, 26.3 keV.
     n_channels
         Number of channels.
 
@@ -165,27 +165,24 @@ def calibrate_from_am241(
         f.read(128)  # skip header
         counts = np.array(struct.unpack(f"{n_channels}i", f.read(n_channels * 4)), dtype=np.int64)
 
-    # Exclude endpoint region (last 200 channels)
+    # Search for peaks in low-energy region only
-    counts_no_end = counts.copy()
+    counts_low = counts.copy()
-    counts_no_end[-200:] = 0
+    counts_low[300:] = 0
-
+    peaks, props = find_peaks(counts_low, height=10, distance=10)
-    peaks, props = find_peaks(counts_no_end, height=10, distance=10)
+    peak_energies = np.array(peaks, dtype=float)
-    peak_energies = np.array(peaks)
     peak_heights = props["peak_heights"]
 
     # Known Am-241 gamma lines (keV)
-    am241_lines = {13.8: None, 17.8: None, 26.3: None, 59.54: None}
+    am241_lines = {13.8: None, 17.8: None, 26.3: None}
     if known_peaks:
         am241_lines = known_peaks
     else:
-        # Find peaks near expected channels
+        # Match lines to peaks by proximity
-        for E, ch in list(am241_lines.items()):
+        for E in am241_lines:
-            if ch is None:
+            best = peak_energies[np.argmin(np.abs(peak_energies - E / 0.1))]
-                # Find the nearest peak
-                best = peak_energies[np.argmin(np.abs(peak_energies - E / 0.45))]  # rough guess
             am241_lines[E] = int(round(best))
 
-    # Use all identified peaks for calibration
+    # Use all matched peaks for calibration
     energies = np.array(list(am241_lines.keys()))
     channels = np.array(list(am241_lines.values()))
 

scripts/analyze_raw.py

@@ -29,11 +29,12 @@ from exp_data.raw_reader import (
 )
 
 # Energy calibration from Am-241 source: E = offset + gain * channel
-# Determined automatically from known Am-241 gamma lines
+# Determined automatically from known Am-241 gamma lines (13.8, 17.8, 26.3 keV)
 CALIBRATION_FILE = "data/raw/CAM1.DAT"
 _calib_offset, _calib_gain = calibrate_from_am241(CALIBRATION_FILE)
 ENERGY_OFFSET = _calib_offset
 ENERGY_GAIN = _calib_gain
+print(f"Energy calibration: E = {ENERGY_OFFSET:.4f} + {ENERGY_GAIN:.4f} * ch")
 
 # Approximate Tc-99 endpoint for reference
 TC99_ENDPOINT = 294.0  # keV
@@ -56,13 +57,13 @@ def estimate_endpoint_from_counts(
     We use a simplified Kurie plot (sqrt(counts)) since the phase space
     factor is slowly varying compared to the sharp drop at the endpoint.
 
-    Fits a line to the last ~10% of the spectrum (in the tail region
+    Fits a line to the last ~15% of the spectrum (in the tail region
     where counts are small but above background) and finds the x-intercept.
 
     Parameters
     ----------
     counts : np.ndarray
-        Count spectrum (channel indices, not energies).
+        Count spectrum.
     energies_keV : np.ndarray
         Energy calibration for each channel.
     guess_endpoint : float
@@ -77,17 +78,18 @@ def estimate_endpoint_from_counts(
     counts_work = counts[:-1]
     energies_work = energies_keV[:-1]
 
-    # Estimate background from the very tail (last 3%)
+    # Estimate background from the very tail (last 5%)
-    n_tail = max(10, int(len(counts_work) * 0.03))
+    n_tail = max(20, int(len(counts_work) * 0.05))
     background = np.median(counts_work[-n_tail:])
 
-    # Select tail region: last 10% of spectrum, above background
+    # Select tail region: last 15% of spectrum, above background
-    n_tail_region = max(20, int(len(counts_work) * 0.10))
+    n_tail_region = max(30, int(len(counts_work) * 0.15))
     E_tail = energies_work[-n_tail_region:]
     c_tail = counts_work[-n_tail_region:]
 
     # Only keep channels above background + noise
-    mask = c_tail > (background + np.sqrt(background))
+    noise = max(np.sqrt(background), 1.0)
+    mask = c_tail > (background + noise)
     if not mask.any():
         return guess_endpoint
 
@@ -99,11 +101,11 @@ def estimate_endpoint_from_counts(
 
     # Linear fit: K = a + b * E, endpoint = -a/b
     try:
-        slope, intercept, _, _ = linregress(E_sel, K)[:4]
+        slope, intercept, r_val, _, _ = linregress(E_sel, K)[:5]
-        if slope < 0:
+        if slope < 0 and abs(r_val) > 0.3:
             endpoint = -intercept / slope
-            # Clamp to reasonable range
+            # Clamp to reasonable range (Q ± 100 keV)
-            endpoint = max(guess_endpoint - 30, min(guess_endpoint + 80, endpoint))
+            endpoint = max(guess_endpoint - 100, min(guess_endpoint + 100, endpoint))
             return float(endpoint)
     except (ValueError, ZeroDivisionError):
         pass